[llvm] [AMDGPU][True16][CodeGen] stop emitting spgr_lo16 from isel (PR #144819)

Brox Chen via llvm-commits llvm-commits at lists.llvm.org
Fri Jun 20 12:14:41 PDT 2025


https://github.com/broxigarchen updated https://github.com/llvm/llvm-project/pull/144819

>From 56ed964c7e347b051b9ba5a33bc2d088f211b949 Mon Sep 17 00:00:00 2001
From: guochen2 <guochen2 at amd.com>
Date: Wed, 18 Jun 2025 21:19:43 -0400
Subject: [PATCH] support -O0 flow intrue16

---
 llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp    |     2 +-
 llvm/lib/Target/AMDGPU/SIInstrInfo.cpp        |    14 +-
 llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp     |     4 +-
 llvm/lib/Target/AMDGPU/VOP1Instructions.td    |    23 +-
 .../CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll  | 45159 ++++++----------
 .../CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll   |  1727 +-
 .../CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll    |   300 +-
 .../CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll   |  6995 +--
 .../CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll   |  8333 +--
 .../CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll    |   589 +-
 .../CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll    |   312 +-
 .../CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll   | 13755 ++---
 .../CodeGen/AMDGPU/amdgcn.bitcast.576bit.ll   |  6882 +--
 .../CodeGen/AMDGPU/amdgcn.bitcast.640bit.ll   |  5262 +-
 .../CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll    |  2112 +-
 .../CodeGen/AMDGPU/amdgcn.bitcast.704bit.ll   |  5812 +-
 .../CodeGen/AMDGPU/amdgcn.bitcast.768bit.ll   |  6376 +--
 .../CodeGen/AMDGPU/amdgcn.bitcast.832bit.ll   |  6992 +--
 .../CodeGen/AMDGPU/amdgcn.bitcast.896bit.ll   |  7496 +--
 .../CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll   |  8132 +--
 .../CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll    |  1055 +-
 .../atomic_optimizations_global_pointer.ll    |    78 +-
 llvm/test/CodeGen/AMDGPU/br_cc.f16.ll         |    35 +-
 llvm/test/CodeGen/AMDGPU/bswap.ll             |   145 +-
 .../CodeGen/AMDGPU/call-argument-types.ll     |    94 +-
 llvm/test/CodeGen/AMDGPU/fabs.bf16.ll         |   118 +-
 llvm/test/CodeGen/AMDGPU/fcopysign.bf16.ll    |    81 +-
 llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll     |   216 +-
 llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll    |   363 +-
 llvm/test/CodeGen/AMDGPU/frem.ll              |   495 +-
 llvm/test/CodeGen/AMDGPU/function-returns.ll  |    79 +-
 .../AMDGPU/gfx-callable-argument-types.ll     |   731 +-
 .../CodeGen/AMDGPU/insert-delay-alu-bug.ll    |   466 +-
 .../CodeGen/AMDGPU/insert_vector_elt.v2i16.ll |   198 +-
 .../AMDGPU/llvm.amdgcn.intersect_ray.ll       |   271 +-
 .../llvm.amdgcn.raw.ptr.tbuffer.load.d16.ll   |    23 +-
 .../llvm.amdgcn.raw.tbuffer.load.d16.ll       |    48 +-
 ...llvm.amdgcn.struct.ptr.tbuffer.load.d16.ll |    26 +-
 .../llvm.amdgcn.struct.tbuffer.load.d16.ll    |    54 +-
 llvm/test/CodeGen/AMDGPU/load-constant-i16.ll |   105 +-
 llvm/test/CodeGen/AMDGPU/load-constant-i8.ll  |   145 +-
 llvm/test/CodeGen/AMDGPU/select.f16.ll        |    67 +-
 42 files changed, 46131 insertions(+), 85039 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp b/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
index 1bf5b4a241780..555455d704c84 100644
--- a/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
@@ -1088,7 +1088,7 @@ void SIFixSGPRCopies::lowerVGPR2SGPRCopies(MachineFunction &MF) {
       assert(MF.getSubtarget<GCNSubtarget>().useRealTrue16Insts() &&
              "We do not expect to see 16-bit copies from VGPR to SGPR unless "
              "we have 16-bit VGPRs");
-      assert(MRI->getRegClass(DstReg) == &AMDGPU::SGPR_LO16RegClass ||
+      assert(MRI->getRegClass(DstReg) == &AMDGPU::SGPR_32RegClass ||
              MRI->getRegClass(DstReg) == &AMDGPU::SReg_32RegClass ||
              MRI->getRegClass(DstReg) == &AMDGPU::SReg_32_XM0RegClass);
       // There is no V_READFIRSTLANE_B16, so legalize the dst/src reg to 32 bits
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 2ebf8b99e9d7b..198695fc9f030 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -7245,7 +7245,8 @@ void SIInstrInfo::legalizeOperandsVALUt16(MachineInstr &MI, unsigned OpIdx,
   MachineBasicBlock *MBB = MI.getParent();
   // Legalize operands and check for size mismatch
   if (!OpIdx || OpIdx >= MI.getNumExplicitOperands() ||
-      OpIdx >= get(Opcode).getNumOperands())
+      OpIdx >= get(Opcode).getNumOperands() ||
+      get(Opcode).operands()[OpIdx].RegClass == -1)
     return;
 
   MachineOperand &Op = MI.getOperand(OpIdx);
@@ -7803,8 +7804,9 @@ void SIInstrInfo::moveToVALUImpl(SIInstrWorklist &Worklist,
       // that copies will end up as machine instructions and not be
       // eliminated.
       addUsersToMoveToVALUWorklist(DstReg, MRI, Worklist);
-      MRI.replaceRegWith(DstReg, Inst.getOperand(1).getReg());
-      MRI.clearKillFlags(Inst.getOperand(1).getReg());
+      Register NewDstReg = Inst.getOperand(1).getReg();
+      MRI.replaceRegWith(DstReg, NewDstReg);
+      MRI.clearKillFlags(NewDstReg);
       Inst.getOperand(0).setReg(DstReg);
       // Make sure we don't leave around a dead VGPR->SGPR copy. Normally
       // these are deleted later, but at -O0 it would leave a suspicious
@@ -7812,6 +7814,12 @@ void SIInstrInfo::moveToVALUImpl(SIInstrWorklist &Worklist,
       for (unsigned I = Inst.getNumOperands() - 1; I != 0; --I)
         Inst.removeOperand(I);
       Inst.setDesc(get(AMDGPU::IMPLICIT_DEF));
+      // Legalize t16 operand since replaceReg is called after addUsersToVALU
+      for (MachineRegisterInfo::use_iterator I = MRI.use_begin(NewDstReg),
+                                             E = MRI.use_end();
+           I != E; ++I) {
+        legalizeOperandsVALUt16(*I->getParent(), MRI);
+      }
       return;
     }
 
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
index e41189adfb46f..8062d1bfe3d20 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -3553,9 +3553,7 @@ SIRegisterInfo::getVectorSuperClassForBitWidth(unsigned BitWidth) const {
 
 const TargetRegisterClass *
 SIRegisterInfo::getSGPRClassForBitWidth(unsigned BitWidth) {
-  if (BitWidth == 16)
-    return &AMDGPU::SGPR_LO16RegClass;
-  if (BitWidth == 32)
+  if (BitWidth == 16 || BitWidth == 32)
     return &AMDGPU::SReg_32RegClass;
   if (BitWidth == 64)
     return &AMDGPU::SReg_64RegClass;
diff --git a/llvm/lib/Target/AMDGPU/VOP1Instructions.td b/llvm/lib/Target/AMDGPU/VOP1Instructions.td
index 926df955881e0..ac6725ded4f10 100644
--- a/llvm/lib/Target/AMDGPU/VOP1Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP1Instructions.td
@@ -1470,17 +1470,34 @@ def : GCNPat<
 >;
 
 def : GCNPat<
-  (i64 (anyext i16:$src)),
+  (i64 (UniformUnaryFrag<anyext> i16:$src)),
+  (REG_SEQUENCE VReg_64,
+     (i32 (COPY $src)), sub0,
+     (V_MOV_B32_e32 (i32 0)), sub1)
+>;
+
+def : GCNPat<
+  (i64 (DivergentUnaryFrag<anyext> i16:$src)),
   (REG_SEQUENCE VReg_64, $src, lo16, (i16 (IMPLICIT_DEF)), hi16, (i32 (IMPLICIT_DEF)), sub1)
 >;
 
 def : GCNPat<
-  (i16 (trunc i32:$src)),
+  (i16 (UniformUnaryFrag<trunc> i32:$src)),
+  (COPY $src)
+>;
+
+def : GCNPat<
+  (i16 (DivergentUnaryFrag<trunc> i32:$src)),
   (EXTRACT_SUBREG $src, lo16)
 >;
 
 def : GCNPat <
-  (i16 (trunc i64:$src)),
+  (i16 (UniformUnaryFrag<trunc> i64:$src)),
+  (EXTRACT_SUBREG $src, sub0)
+>;
+
+def : GCNPat <
+  (i16 (DivergentUnaryFrag<trunc> i64:$src)),
   (EXTRACT_SUBREG $src, lo16)
 >;
 
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
index f729de82cb042..cb2f0f28a29d6 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
@@ -10438,1974 +10438,951 @@ define inreg <128 x i8> @bitcast_v32i32_to_v128i8_scalar(<32 x i32> inreg %a, i3
 ; GFX9-NEXT:    ; kill: killed $sgpr47
 ; GFX9-NEXT:    s_branch .LBB13_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32i32_to_v128i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v16, s32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v17, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v18, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v19, s32 offset:12
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s30, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s96, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s97, 1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s98, 2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s99, 3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s100, 4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s101, 5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr18 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr19 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s102, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s103, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s104, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s50, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s51, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s52, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s53, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s54, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s55, 15
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s64, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s65, 17
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s66, 18
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s67, 19
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s68, 20
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s69, 21
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s70, 22
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s71, 23
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s80, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s81, 25
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s82, 26
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s83, 27
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s84, 28
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s85, 29
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s86, 30
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s87, 31
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[4:5], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s100, s15, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s101, s40, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s29, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s29, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 31
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s4, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s28, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 30
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s4, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[6:7], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s27, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 29
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s27, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 28
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s26, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 27
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s25, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 26
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s6, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[10:11], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s25, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 25
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s6, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s51, s24, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s52, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s53, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s54, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 23
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s55, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s64, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s65, s22, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 22
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s66, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s67, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s68, s21, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 21
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s69, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s70, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s71, s19, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 20
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s80, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s81, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s82, s18, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 19
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s83, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s84, s17, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 18
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s85, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 17
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s86, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s87, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s96, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s97, s1, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 15
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s98, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s0, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 14
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[12:13], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[60:61], s[14:15], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 13
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[72:73], s[28:29], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[26:27], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[78:79], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[76:77], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[88:89], s[20:21], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 11
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s12, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s14, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s14, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s40, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s29, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s18, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s57
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-TRUE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 24
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 31
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s4, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s13, s13, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 30
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s4, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s15, s15, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[4:5], 24
-; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 29
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s40, s40, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 28
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[6:7], 24
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 27
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 6
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 26
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s6, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 25
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s6, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 4
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[10:11], 24
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 23
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 2
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 22
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 21
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 20
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 19
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s100, s15, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s101, s40, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s29, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 18
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s29, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s28, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 17
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s27, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s25, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 15
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s51, s24, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 14
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s52, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s53, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s54, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 13
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s55, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s64, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s65, s22, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s66, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s67, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s68, s21, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 11
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s69, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s70, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s71, s19, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s12, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s80, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s81, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s82, s18, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s83, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s84, s17, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s85, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s86, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s87, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s14, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s96, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s97, s1, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s98, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s0, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[12:13], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[60:61], s[14:15], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[72:73], s[28:29], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[26:27], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s40, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[78:79], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[76:77], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[88:89], s[20:21], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s29, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
-; GFX11-TRUE16-NEXT:  .LBB13_3: ; %end
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s56, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s99
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s30
-; GFX11-TRUE16-NEXT:    s_and_b32 s57, s57, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s56, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_or_b32 s56, s57, s56
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s47, s47, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s56, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s47
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s94
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s98
-; GFX11-TRUE16-NEXT:    s_and_b32 s46, s46, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s47, s47, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s56, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s46, s46, s47
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s97
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s96
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s46, s46, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s56, s56, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s57, s57, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s46
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s87
-; GFX11-TRUE16-NEXT:    s_or_b32 s56, s56, s57
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s46, s46, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s56, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s46
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s86
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s56
-; GFX11-TRUE16-NEXT:    s_and_b32 s46, s46, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s45, s45, 8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s0 :: v_dual_mov_b32 v4, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s45, s46, s45
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s44
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s45, s45, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s45
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s2 :: v_dual_mov_b32 v6, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s92
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s85
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s84
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s43
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s83
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s82
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s18
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s42
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s71
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s90
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s81
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s16, s16, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s0 :: v_dual_mov_b32 v8, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s70
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s80
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s16, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s88
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s69
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s68
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s66
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s67
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s64
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s22
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s65
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s55
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s76
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s53
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s23
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s18, s18, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s54
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s17, s17, 0xff
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s17, s17, s18
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s16, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s16, s17, 16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[3:6], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s52
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s51
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s78
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s50
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s49
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s48
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s39
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s38
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s74
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s36
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s35
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s18, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s19, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-TRUE16-NEXT:    s_or_b32 s17, s18, s19
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s104
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s72
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s17, v18, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s103
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s102
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s16, v18, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s101
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s62
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s17, v18, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s18, v18, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s19, v18, 4
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s41
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s18, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s19, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-TRUE16-NEXT:    s_or_b32 s17, s18, s19
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s1, v18, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s2, v18, 6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s60
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s14, v18, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s100
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, s14, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s15, s15, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s14, s15
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s12, v18, 10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s58
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s14, s14, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, s12, 0xff
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s12, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s13, v18, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s14, v18, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s15, v18, 13
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, s12, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s13, s13, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, s14, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s15, s15, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s12, s12, s13
-; GFX11-TRUE16-NEXT:    s_or_b32 s13, s14, s15
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, s12, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s13, s13, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s12, s13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s12, v19, 2
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s1, v18, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s2, v18, 15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s12
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 16
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s10, v18, 17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s11, v18, 18
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s10, s11
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 19
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s8, v18, 20
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s10, v19, 4
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s11, v19, 5
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s10, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s8, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s9, v18, 21
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s10, v18, 22
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s11, v18, 23
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s8, s9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s8, v19, 6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s1, v18, 24
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s2, v18, 25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 26
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s6, v18, 27
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s7, v18, 28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s6, s7
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 29
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s4, v18, 30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s6, v19, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s7, v19, 9
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s5, v18, 31
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s6, v19, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s7, v19, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s13, v19, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s9, v19, 7
-; GFX11-TRUE16-NEXT:    s_clause 0x2
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:112
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s104, v17, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s103, v17, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s102, v17, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s101, v17, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s100, v17, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s99, v17, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s98, v17, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s97, v17, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s96, v17, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s87, v16, 31
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s86, v16, 30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s85, v16, 29
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s84, v16, 28
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s83, v16, 27
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s82, v16, 26
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s81, v16, 25
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s80, v16, 24
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s71, v16, 23
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s70, v16, 22
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s69, v16, 21
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s68, v16, 20
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v16, 19
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v16, 18
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v16, 17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v16, 16
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v16, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v16, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v16, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v16, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v16, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v16, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v16, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v16, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v16, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v16, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v16, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v16, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v16, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v16, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v16, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v16, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v16, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v17, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v18, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v19, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB13_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr101_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr99_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr98_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr97_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr96_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr87_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr86_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr85_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr84_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr104_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr103_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr102_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 2
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 8
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 9
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 10
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 11
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 12
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 13
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 14
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 15
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 17
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 18
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 19
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 20
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 21
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 22
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 23
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 24
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 25
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 26
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 27
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 28
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 29
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 30
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s100, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s101, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s100, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s101, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s100, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s101, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s100, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s101, 9
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB13_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v32i32_to_v128i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v16, s32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v17, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v18, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v19, s32 offset:12
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s96, 0
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s97, 1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s98, 2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s99, 3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s100, 4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s101, 5
-; GFX11-FAKE16-NEXT:    s_mov_b32 s101, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr19 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr18 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s102, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s103, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s104, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s49, 9
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s50, 10
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s51, 11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s52, 12
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s53, 13
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s54, 14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s55, 15
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s64, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s65, 17
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s66, 18
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s67, 19
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s68, 20
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s69, 21
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s70, 22
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s71, 23
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s80, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s81, 25
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s82, 26
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s83, 27
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s84, 28
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s85, 29
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s86, 30
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s87, 31
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB13_2
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[4:5], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s5, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s5, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 17
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s7, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 18
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s7, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s69, s6, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 19
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s6, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s9, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s9, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 20
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s9, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s8, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 21
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s11, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s50, s11, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 22
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s51, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s52, s10, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s53, s13, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 23
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s54, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s55, s13, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s64, s12, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s65, s12, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s66, s15, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s67, s15, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 25
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s68, s15, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s14, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 26
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s70, s41, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s71, s41, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s41, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 27
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s80, s40, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s40, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s81, s29, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 28
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s82, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s83, s29, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s84, s28, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 29
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s85, s28, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s86, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s27, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 30
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s87, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s96, s26, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 31
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s97, s25, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 14
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s2, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s17, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 15
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[6:7], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s98, s1, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s17, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s99, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s1, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s17, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 13
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[8:9], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 10
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[76:77], s[26:27], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 11
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[10:11], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[88:89], s[24:25], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[14:15], 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[28:29], 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_branch .LBB13_3
-; GFX11-FAKE16-NEXT:  .LBB13_2:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    s_mov_b32 s101, -1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 0
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 1
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 2
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 3
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr100
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr99
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr98
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr97
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr96
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr87
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr86
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr85
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr103
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr102
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr104
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 10
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 12
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 15
-; GFX11-FAKE16-NEXT:  .LBB13_3: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s101
-; GFX11-FAKE16-NEXT:    s_mov_b32 s101, s104
-; GFX11-FAKE16-NEXT:    s_mov_b32 s104, s57
-; GFX11-FAKE16-NEXT:    s_mov_b32 s57, s69
-; GFX11-FAKE16-NEXT:    s_mov_b32 s69, s42
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB13_5
-; GFX11-FAKE16-NEXT:  ; %bb.4: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s24, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 17
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s24, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 18
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[4:5], 24
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 19
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 20
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s13, s13, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 21
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s22, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s15, s15, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 22
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s22, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s41, s41, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 23
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 24
-; GFX11-FAKE16-NEXT:    s_add_i32 s40, s40, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 25
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 26
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s101, s5, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s5, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 27
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s7, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 28
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s7, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s6, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 29
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s6, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s9, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s9, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 30
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s9, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s8, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 31
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s11, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 14
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s50, s11, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 15
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[6:7], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s51, s10, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s52, s10, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s53, s13, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 13
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[8:9], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s54, s13, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 10
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s55, s13, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s64, s12, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 11
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[10:11], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s65, s12, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s66, s15, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s67, s15, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s68, s15, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s14, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s14, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s70, s41, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s71, s41, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s41, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[14:15], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s80, s40, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s40, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s81, s29, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s82, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s83, s29, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s84, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s85, s28, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s86, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s87, s27, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s96, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s97, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s69, s25, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[28:29], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s98, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s99, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[76:77], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[88:89], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 9
-; GFX11-FAKE16-NEXT:  .LBB13_5: ; %end
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s43, s43, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, s74, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s43
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s43, s94, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s42, s42, s43
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s45, s45, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s42, s42, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s42
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s42, v18, 9
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s45, s30, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s44, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s44, s44, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s42, s42, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s44, s44, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s42
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s42, v18, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s43, v18, 7
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s44
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s44, s100, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s45, s98, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s44
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s99, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, s42, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s44, s44, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s43, s43, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s44, s44, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s42, s42, s43
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s44
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s42, s42, 16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v18, 6
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s42
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v18, 5
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s3, v18, 2
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s92, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v18, 4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v18, 0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v18, 3
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s17, v19, 29
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s100, v17, 4
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s99, v17, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v18, 1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v19, 28
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s18, v19, 19
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s90, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v19, 31
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s19, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s16, s16, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v19, 30
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s78, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s18, s18, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s19, s86, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s17, v19, 21
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s16, s16, 16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s98, v17, 2
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v19, 27
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s3, v19, 24
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v19, 22
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v19, 26
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s21, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s86, v16, 30
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v19, 25
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v16, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v16, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v19, 23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s1, v19, 18
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s62, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s23, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s17, v19, 20
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s16, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s17, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s17, s17, s18
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s16, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s97, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v19, 17
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s88, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s69, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s18, s72, 0xff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s97, v17, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s69, v16, 21
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s3, v19, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s25, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s73, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s96, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s76, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s27, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s87, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-FAKE16-NEXT:    s_or_b32 s17, s18, s19
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s16, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s17, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v19, 0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s85, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s84, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s16, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s17, v19, 1
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s29, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s83, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s82, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s81, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s18, v19, 2
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s40, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s61, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s80, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s18, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s19, v19, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s41, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s60, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s18, s71, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s19, s70, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-FAKE16-NEXT:    s_or_b32 s17, s18, s19
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s16, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s17, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v19, 4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s14, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s58, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s59, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s16, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s15, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s68, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, s67, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s15, s66, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s14, s15
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s14, v19, 6
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s12, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s65, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, s64, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s14, s14, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s15, v19, 7
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s12, s14
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, s13, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s13, s55, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, s54, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s15, s53, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s12, s12, s13
-; GFX11-FAKE16-NEXT:    s_or_b32 s13, s14, s15
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, s12, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s13, s13, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s12, s13
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s12, v19, 8
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s10, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s52, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s51, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s12, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s11, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s50, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s49, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s48, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s10, s11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s10, v19, 10
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s8, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s39, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s38, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s10, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s11, v19, 11
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s8, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s9, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s37, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s36, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s35, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s8, s9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s8, v19, 12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s6, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s56, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s57, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s8, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s7, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s34, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_hi, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s46, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s6, s7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s6, v19, 14
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s4, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s47, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s104, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s7, v19, 15
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s5, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s103, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s102, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s101, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:64
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s17, v19, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s13, v19, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s9, v19, 13
-; GFX11-FAKE16-NEXT:    s_clause 0x2
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:112
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s104, v17, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s103, v17, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s102, v17, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s101, v17, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s96, v17, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s87, v16, 31
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s85, v16, 29
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s84, v16, 28
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s83, v16, 27
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s82, v16, 26
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s81, v16, 25
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s80, v16, 24
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s71, v16, 23
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s70, v16, 22
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s68, v16, 20
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v16, 19
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v16, 18
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v16, 17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v16, 16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v16, 15
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v16, 14
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v16, 13
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v16, 12
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v16, 11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v16, 10
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v16, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v16, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v16, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v16, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v16, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v16, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v16, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v16, 2
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v16, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v17, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v18, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v19, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v32i32_to_v128i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b32 off, v16, s32
+; GFX11-NEXT:    scratch_store_b32 off, v17, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v18, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v19, s32 offset:12
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v16, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v17, s96, 0
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
+; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s41, v2
+; GFX11-NEXT:    v_writelane_b32 v16, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v17, s97, 1
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s15, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
+; GFX11-NEXT:    v_writelane_b32 v16, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v17, s98, 2
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v8
+; GFX11-NEXT:    v_writelane_b32 v16, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v17, s99, 3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
+; GFX11-NEXT:    v_writelane_b32 v16, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v17, s100, 4
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v14
+; GFX11-NEXT:    v_writelane_b32 v16, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v17, s101, 5
+; GFX11-NEXT:    s_mov_b32 s101, 0
+; GFX11-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
+; GFX11-NEXT:    ; implicit-def: $vgpr19 : SGPR spill to VGPR lane
+; GFX11-NEXT:    ; implicit-def: $vgpr18 : SGPR spill to VGPR lane
+; GFX11-NEXT:    v_writelane_b32 v16, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v17, s102, 6
+; GFX11-NEXT:    v_writelane_b32 v16, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v17, s103, 7
+; GFX11-NEXT:    v_writelane_b32 v16, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v17, s104, 8
+; GFX11-NEXT:    v_writelane_b32 v16, s49, 9
+; GFX11-NEXT:    v_writelane_b32 v16, s50, 10
+; GFX11-NEXT:    v_writelane_b32 v16, s51, 11
+; GFX11-NEXT:    v_writelane_b32 v16, s52, 12
+; GFX11-NEXT:    v_writelane_b32 v16, s53, 13
+; GFX11-NEXT:    v_writelane_b32 v16, s54, 14
+; GFX11-NEXT:    v_writelane_b32 v16, s55, 15
+; GFX11-NEXT:    v_writelane_b32 v16, s64, 16
+; GFX11-NEXT:    v_writelane_b32 v16, s65, 17
+; GFX11-NEXT:    v_writelane_b32 v16, s66, 18
+; GFX11-NEXT:    v_writelane_b32 v16, s67, 19
+; GFX11-NEXT:    v_writelane_b32 v16, s68, 20
+; GFX11-NEXT:    v_writelane_b32 v16, s69, 21
+; GFX11-NEXT:    v_writelane_b32 v16, s70, 22
+; GFX11-NEXT:    v_writelane_b32 v16, s71, 23
+; GFX11-NEXT:    v_writelane_b32 v16, s80, 24
+; GFX11-NEXT:    v_writelane_b32 v16, s81, 25
+; GFX11-NEXT:    v_writelane_b32 v16, s82, 26
+; GFX11-NEXT:    v_writelane_b32 v16, s83, 27
+; GFX11-NEXT:    v_writelane_b32 v16, s84, 28
+; GFX11-NEXT:    v_writelane_b32 v16, s85, 29
+; GFX11-NEXT:    v_writelane_b32 v16, s86, 30
+; GFX11-NEXT:    v_writelane_b32 v16, s87, 31
+; GFX11-NEXT:    s_cbranch_scc0 .LBB13_2
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s43, s25, 8
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[4:5], 24
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s104, s5, 24
+; GFX11-NEXT:    s_lshr_b32 s102, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s103, s5, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 17
+; GFX11-NEXT:    s_lshr_b32 s43, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s57, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 8
+; GFX11-NEXT:    s_lshr_b32 s46, s7, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 18
+; GFX11-NEXT:    s_lshr_b32 s43, s23, 24
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s34, s7, 8
+; GFX11-NEXT:    s_lshr_b32 s69, s6, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 19
+; GFX11-NEXT:    s_lshr_b32 s43, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s6, 8
+; GFX11-NEXT:    s_lshr_b32 s35, s9, 24
+; GFX11-NEXT:    s_lshr_b32 s36, s9, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 20
+; GFX11-NEXT:    s_lshr_b32 s43, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s37, s9, 8
+; GFX11-NEXT:    s_lshr_b32 s38, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s39, s8, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 21
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s11, 24
+; GFX11-NEXT:    s_lshr_b32 s49, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s50, s11, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 22
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s51, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s52, s10, 8
+; GFX11-NEXT:    s_lshr_b32 s53, s13, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 23
+; GFX11-NEXT:    s_lshr_b32 s43, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s54, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s55, s13, 8
+; GFX11-NEXT:    s_lshr_b32 s64, s12, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s65, s12, 8
+; GFX11-NEXT:    s_lshr_b32 s66, s15, 24
+; GFX11-NEXT:    s_lshr_b32 s67, s15, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 25
+; GFX11-NEXT:    s_lshr_b32 s43, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s68, s15, 8
+; GFX11-NEXT:    s_lshr_b32 s59, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s14, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 26
+; GFX11-NEXT:    s_lshr_b32 s43, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s70, s41, 24
+; GFX11-NEXT:    s_lshr_b32 s71, s41, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s41, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 27
+; GFX11-NEXT:    s_lshr_b32 s43, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s80, s40, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s40, 8
+; GFX11-NEXT:    s_lshr_b32 s81, s29, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 28
+; GFX11-NEXT:    s_lshr_b32 s43, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s82, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s83, s29, 8
+; GFX11-NEXT:    s_lshr_b32 s84, s28, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 29
+; GFX11-NEXT:    s_lshr_b32 s43, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s85, s28, 8
+; GFX11-NEXT:    s_lshr_b32 s86, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s72, s27, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 30
+; GFX11-NEXT:    s_lshr_b32 s43, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s87, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s73, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s96, s26, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 31
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s97, s25, 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 0
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 14
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s2, 16
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 1
+; GFX11-NEXT:    s_lshr_b32 s43, s17, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 15
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[6:7], 24
+; GFX11-NEXT:    s_lshr_b32 s98, s1, 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 2
+; GFX11-NEXT:    s_lshr_b32 s43, s17, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 12
+; GFX11-NEXT:    s_lshr_b32 s99, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s100, s1, 8
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 3
+; GFX11-NEXT:    s_lshr_b32 s43, s17, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 13
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[8:9], 24
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 4
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 10
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[76:77], s[26:27], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 5
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 11
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[10:11], 24
+; GFX11-NEXT:    s_lshr_b64 s[88:89], s[24:25], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 6
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 8
+; GFX11-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 7
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 9
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
+; GFX11-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 6
+; GFX11-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 9
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 7
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[14:15], 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 4
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 5
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 2
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 3
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[28:29], 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 0
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 1
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
+; GFX11-NEXT:    s_branch .LBB13_3
+; GFX11-NEXT:  .LBB13_2:
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    s_mov_b32 s101, -1
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 0
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 1
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 2
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 3
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 4
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 5
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $sgpr100
+; GFX11-NEXT:    ; implicit-def: $sgpr99
+; GFX11-NEXT:    ; implicit-def: $sgpr98
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr97
+; GFX11-NEXT:    ; implicit-def: $sgpr96
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr87
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr86
+; GFX11-NEXT:    ; implicit-def: $sgpr85
+; GFX11-NEXT:    ; implicit-def: $sgpr84
+; GFX11-NEXT:    ; implicit-def: $sgpr83
+; GFX11-NEXT:    ; implicit-def: $sgpr82
+; GFX11-NEXT:    ; implicit-def: $sgpr81
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr80
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr71
+; GFX11-NEXT:    ; implicit-def: $sgpr70
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr68
+; GFX11-NEXT:    ; implicit-def: $sgpr67
+; GFX11-NEXT:    ; implicit-def: $sgpr66
+; GFX11-NEXT:    ; implicit-def: $sgpr65
+; GFX11-NEXT:    ; implicit-def: $sgpr64
+; GFX11-NEXT:    ; implicit-def: $sgpr55
+; GFX11-NEXT:    ; implicit-def: $sgpr54
+; GFX11-NEXT:    ; implicit-def: $sgpr53
+; GFX11-NEXT:    ; implicit-def: $sgpr52
+; GFX11-NEXT:    ; implicit-def: $sgpr51
+; GFX11-NEXT:    ; implicit-def: $sgpr50
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr69
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr103
+; GFX11-NEXT:    ; implicit-def: $sgpr102
+; GFX11-NEXT:    ; implicit-def: $sgpr104
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 6
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 7
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 8
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 9
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 10
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 11
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 12
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 13
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 14
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 15
+; GFX11-NEXT:  .LBB13_3: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s101
+; GFX11-NEXT:    s_mov_b32 s101, s104
+; GFX11-NEXT:    s_mov_b32 s104, s57
+; GFX11-NEXT:    s_mov_b32 s57, s69
+; GFX11-NEXT:    s_mov_b32 s69, s42
+; GFX11-NEXT:    s_cbranch_vccnz .LBB13_5
+; GFX11-NEXT:  ; %bb.4: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 8
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s24, 16
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 17
+; GFX11-NEXT:    s_lshr_b32 s42, s24, 8
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 18
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 24
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[4:5], 24
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 19
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_add_i32 s9, s9, 3
+; GFX11-NEXT:    s_add_i32 s8, s8, 3
+; GFX11-NEXT:    s_add_i32 s11, s11, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 20
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 8
+; GFX11-NEXT:    s_add_i32 s10, s10, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s13, s13, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 21
+; GFX11-NEXT:    s_lshr_b32 s42, s22, 16
+; GFX11-NEXT:    s_add_i32 s12, s12, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s15, s15, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 22
+; GFX11-NEXT:    s_lshr_b32 s42, s22, 8
+; GFX11-NEXT:    s_add_i32 s14, s14, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s41, s41, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 23
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 24
+; GFX11-NEXT:    s_add_i32 s40, s40, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 16
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 25
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 8
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 26
+; GFX11-NEXT:    s_lshr_b32 s42, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s101, s5, 24
+; GFX11-NEXT:    s_lshr_b32 s102, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s103, s5, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 27
+; GFX11-NEXT:    s_lshr_b32 s42, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s104, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 8
+; GFX11-NEXT:    s_lshr_b32 s46, s7, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 28
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 24
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s34, s7, 8
+; GFX11-NEXT:    s_lshr_b32 s57, s6, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 29
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s6, 8
+; GFX11-NEXT:    s_lshr_b32 s35, s9, 24
+; GFX11-NEXT:    s_lshr_b32 s36, s9, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 30
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s37, s9, 8
+; GFX11-NEXT:    s_lshr_b32 s38, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s39, s8, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 31
+; GFX11-NEXT:    s_lshr_b32 s42, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s11, 24
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 0
+; GFX11-NEXT:    s_lshr_b32 s42, s18, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 14
+; GFX11-NEXT:    s_lshr_b32 s49, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s50, s11, 8
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 1
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 15
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[6:7], 24
+; GFX11-NEXT:    s_lshr_b32 s51, s10, 16
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 2
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 12
+; GFX11-NEXT:    s_lshr_b32 s52, s10, 8
+; GFX11-NEXT:    s_lshr_b32 s53, s13, 24
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 13
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[8:9], 24
+; GFX11-NEXT:    s_lshr_b32 s54, s13, 16
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 4
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 10
+; GFX11-NEXT:    s_lshr_b32 s55, s13, 8
+; GFX11-NEXT:    s_lshr_b32 s64, s12, 16
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 5
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 11
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[10:11], 24
+; GFX11-NEXT:    s_lshr_b32 s65, s12, 8
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 6
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 8
+; GFX11-NEXT:    s_lshr_b32 s66, s15, 24
+; GFX11-NEXT:    s_lshr_b32 s67, s15, 16
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 7
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 9
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
+; GFX11-NEXT:    s_lshr_b32 s68, s15, 8
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 8
+; GFX11-NEXT:    s_lshr_b32 s59, s14, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 6
+; GFX11-NEXT:    s_lshr_b32 s58, s14, 8
+; GFX11-NEXT:    s_lshr_b32 s70, s41, 24
+; GFX11-NEXT:    s_lshr_b32 s71, s41, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s41, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 7
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[14:15], 24
+; GFX11-NEXT:    s_lshr_b32 s80, s40, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s40, 8
+; GFX11-NEXT:    s_lshr_b32 s81, s29, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 4
+; GFX11-NEXT:    s_lshr_b32 s82, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s83, s29, 8
+; GFX11-NEXT:    s_lshr_b32 s84, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s85, s28, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 5
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
+; GFX11-NEXT:    s_lshr_b32 s86, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s72, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s87, s27, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 2
+; GFX11-NEXT:    s_lshr_b32 s73, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s96, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s97, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s69, s25, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 3
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[28:29], 24
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s74, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 0
+; GFX11-NEXT:    s_lshr_b32 s98, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s99, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s100, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 1
+; GFX11-NEXT:    s_lshr_b64 s[76:77], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[88:89], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 9
+; GFX11-NEXT:  .LBB13_5: ; %end
+; GFX11-NEXT:    s_lshl_b32 s43, s43, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_and_b32 s42, s74, 0xff
+; GFX11-NEXT:    s_or_b32 s2, s2, s43
+; GFX11-NEXT:    s_lshl_b32 s43, s94, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_or_b32 s42, s42, s43
+; GFX11-NEXT:    s_lshl_b32 s45, s45, 8
+; GFX11-NEXT:    s_lshl_b32 s42, s42, 16
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_or_b32 s2, s2, s42
+; GFX11-NEXT:    v_readlane_b32 s42, v18, 9
+; GFX11-NEXT:    s_or_b32 s0, s0, s45
+; GFX11-NEXT:    s_lshl_b32 s45, s30, 8
+; GFX11-NEXT:    s_and_b32 s44, s44, 0xff
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xff
+; GFX11-NEXT:    s_or_b32 s44, s44, s45
+; GFX11-NEXT:    s_lshl_b32 s42, s42, 8
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s44, s44, 16
+; GFX11-NEXT:    s_or_b32 s3, s3, s42
+; GFX11-NEXT:    v_readlane_b32 s42, v18, 8
+; GFX11-NEXT:    v_readlane_b32 s43, v18, 7
+; GFX11-NEXT:    s_or_b32 s0, s0, s44
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xff
+; GFX11-NEXT:    s_lshl_b32 s44, s100, 8
+; GFX11-NEXT:    s_lshl_b32 s45, s98, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s44
+; GFX11-NEXT:    s_and_b32 s44, s99, 0xff
+; GFX11-NEXT:    s_and_b32 s42, s42, 0xff
+; GFX11-NEXT:    s_or_b32 s44, s44, s45
+; GFX11-NEXT:    s_lshl_b32 s43, s43, 8
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s44, s44, 16
+; GFX11-NEXT:    s_or_b32 s42, s42, s43
+; GFX11-NEXT:    s_or_b32 s1, s1, s44
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s42, s42, 16
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_readlane_b32 s0, v18, 6
+; GFX11-NEXT:    s_or_b32 s3, s3, s42
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    v_readlane_b32 s2, v18, 5
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 8
+; GFX11-NEXT:    s_and_b32 s1, s16, 0xff
+; GFX11-NEXT:    v_readlane_b32 s3, v18, 2
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_lshl_b32 s1, s92, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    v_readlane_b32 s2, v18, 4
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s17, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 8
+; GFX11-NEXT:    v_readlane_b32 s16, v18, 0
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_readlane_b32 s2, v18, 3
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    v_readlane_b32 s17, v19, 29
+; GFX11-NEXT:    s_and_b32 s16, s16, 0xff
+; GFX11-NEXT:    v_readlane_b32 s100, v17, 4
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    v_readlane_b32 s99, v17, 3
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_and_b32 s3, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_lshl_b32 s17, s17, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_readlane_b32 s2, v18, 1
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_readlane_b32 s0, v19, 28
+; GFX11-NEXT:    s_and_b32 s1, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 8
+; GFX11-NEXT:    v_readlane_b32 s18, v19, 19
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_lshl_b32 s3, s90, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_or_b32 s3, s16, s3
+; GFX11-NEXT:    v_readlane_b32 s16, v19, 31
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_and_b32 s3, s19, 0xff
+; GFX11-NEXT:    s_lshl_b32 s16, s16, 8
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s3, s3, s16
+; GFX11-NEXT:    v_readlane_b32 s16, v19, 30
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s78, 8
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s18, s18, 8
+; GFX11-NEXT:    s_and_b32 s16, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s19, s86, 8
+; GFX11-NEXT:    s_or_b32 s16, s16, s17
+; GFX11-NEXT:    v_readlane_b32 s17, v19, 21
+; GFX11-NEXT:    s_lshl_b32 s16, s16, 16
+; GFX11-NEXT:    v_readlane_b32 s98, v17, 2
+; GFX11-NEXT:    s_or_b32 s3, s3, s16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    v_readlane_b32 s2, v19, 27
+; GFX11-NEXT:    v_readlane_b32 s3, v19, 24
+; GFX11-NEXT:    v_readlane_b32 s16, v19, 22
+; GFX11-NEXT:    s_lshl_b32 s17, s17, 8
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    v_readlane_b32 s2, v19, 26
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s16, s16, 0xff
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s21, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 8
+; GFX11-NEXT:    v_readlane_b32 s86, v16, 30
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_readlane_b32 s2, v19, 25
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    v_readlane_b32 s31, v16, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v16, 0
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_and_b32 s3, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_readlane_b32 s2, v19, 23
+; GFX11-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
+; GFX11-NEXT:    v_readlane_b32 s1, v19, 18
+; GFX11-NEXT:    s_and_b32 s0, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_lshl_b32 s3, s62, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_or_b32 s3, s16, s3
+; GFX11-NEXT:    s_and_b32 s16, s23, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s16, s16, s17
+; GFX11-NEXT:    v_readlane_b32 s17, v19, 20
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_and_b32 s3, s16, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s17, s17, 0xff
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s17, s17, s18
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s16, s17, 16
+; GFX11-NEXT:    s_lshl_b32 s17, s97, 8
+; GFX11-NEXT:    s_or_b32 s3, s3, s16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
+; GFX11-NEXT:    v_readlane_b32 s2, v19, 17
+; GFX11-NEXT:    s_lshl_b32 s3, s88, 8
+; GFX11-NEXT:    s_and_b32 s16, s69, 0xff
+; GFX11-NEXT:    s_and_b32 s18, s72, 0xff
+; GFX11-NEXT:    v_readlane_b32 s97, v17, 1
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    v_readlane_b32 s69, v16, 21
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    v_readlane_b32 s3, v19, 16
+; GFX11-NEXT:    s_and_b32 s2, s25, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s16, s73, 0xff
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s96, 8
+; GFX11-NEXT:    s_lshl_b32 s17, s76, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    s_and_b32 s16, s27, 0xff
+; GFX11-NEXT:    s_lshl_b32 s17, s87, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_or_b32 s16, s16, s17
+; GFX11-NEXT:    s_or_b32 s17, s18, s19
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s16, s16, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s17, s17, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    v_readlane_b32 s16, v19, 0
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    s_and_b32 s0, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s85, 8
+; GFX11-NEXT:    s_and_b32 s2, s84, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s16, 8
+; GFX11-NEXT:    v_readlane_b32 s17, v19, 1
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s29, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s83, 8
+; GFX11-NEXT:    s_and_b32 s16, s82, 0xff
+; GFX11-NEXT:    s_lshl_b32 s17, s81, 8
+; GFX11-NEXT:    v_readlane_b32 s18, v19, 2
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s40, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s61, 8
+; GFX11-NEXT:    s_and_b32 s16, s80, 0xff
+; GFX11-NEXT:    s_lshl_b32 s17, s18, 8
+; GFX11-NEXT:    v_readlane_b32 s19, v19, 3
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    s_and_b32 s16, s41, 0xff
+; GFX11-NEXT:    s_lshl_b32 s17, s60, 8
+; GFX11-NEXT:    s_and_b32 s18, s71, 0xff
+; GFX11-NEXT:    s_lshl_b32 s19, s70, 8
+; GFX11-NEXT:    s_or_b32 s16, s16, s17
+; GFX11-NEXT:    s_or_b32 s17, s18, s19
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s16, s16, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s17, s17, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    v_readlane_b32 s16, v19, 4
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    s_and_b32 s0, s14, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s58, 8
+; GFX11-NEXT:    s_and_b32 s2, s59, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s16, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s15, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s68, 8
+; GFX11-NEXT:    s_and_b32 s14, s67, 0xff
+; GFX11-NEXT:    s_lshl_b32 s15, s66, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s14, s15
+; GFX11-NEXT:    v_readlane_b32 s14, v19, 6
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s12, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s65, 8
+; GFX11-NEXT:    s_and_b32 s12, s64, 0xff
+; GFX11-NEXT:    s_lshl_b32 s14, s14, 8
+; GFX11-NEXT:    v_readlane_b32 s15, v19, 7
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s12, s14
+; GFX11-NEXT:    s_and_b32 s12, s13, 0xff
+; GFX11-NEXT:    s_lshl_b32 s13, s55, 8
+; GFX11-NEXT:    s_and_b32 s14, s54, 0xff
+; GFX11-NEXT:    s_lshl_b32 s15, s53, 8
+; GFX11-NEXT:    s_or_b32 s12, s12, s13
+; GFX11-NEXT:    s_or_b32 s13, s14, s15
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s12, s12, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s13, s13, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s12, s13
+; GFX11-NEXT:    v_readlane_b32 s12, v19, 8
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
+; GFX11-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
+; GFX11-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
+; GFX11-NEXT:    s_and_b32 s0, s10, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s52, 8
+; GFX11-NEXT:    s_and_b32 s2, s51, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s12, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s11, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s50, 8
+; GFX11-NEXT:    s_and_b32 s10, s49, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s48, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s10, s11
+; GFX11-NEXT:    v_readlane_b32 s10, v19, 10
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s8, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s39, 8
+; GFX11-NEXT:    s_and_b32 s8, s38, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s10, 8
+; GFX11-NEXT:    v_readlane_b32 s11, v19, 11
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s8, s10
+; GFX11-NEXT:    s_and_b32 s8, s9, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s37, 8
+; GFX11-NEXT:    s_and_b32 s10, s36, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s35, 8
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    s_or_b32 s9, s10, s11
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s8, s8, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s9, s9, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s8, s9
+; GFX11-NEXT:    v_readlane_b32 s8, v19, 12
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    s_and_b32 s0, s6, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s56, 8
+; GFX11-NEXT:    s_and_b32 s2, s57, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s8, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s7, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s34, 8
+; GFX11-NEXT:    s_and_b32 s6, vcc_hi, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s46, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s6, s7
+; GFX11-NEXT:    v_readlane_b32 s6, v19, 14
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s4, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s47, 8
+; GFX11-NEXT:    s_and_b32 s4, s104, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 8
+; GFX11-NEXT:    v_readlane_b32 s7, v19, 15
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s6
+; GFX11-NEXT:    s_and_b32 s4, s5, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s103, 8
+; GFX11-NEXT:    s_and_b32 s6, s102, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s101, 8
+; GFX11-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-NEXT:    s_or_b32 s5, s6, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s4, s4, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:64
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    v_readlane_b32 s17, v19, 5
+; GFX11-NEXT:    v_readlane_b32 s13, v19, 9
+; GFX11-NEXT:    v_readlane_b32 s9, v19, 13
+; GFX11-NEXT:    s_clause 0x2
+; GFX11-NEXT:    scratch_store_b128 v0, v[9:12], off offset:80
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:96
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:112
+; GFX11-NEXT:    v_readlane_b32 s104, v17, 8
+; GFX11-NEXT:    v_readlane_b32 s103, v17, 7
+; GFX11-NEXT:    v_readlane_b32 s102, v17, 6
+; GFX11-NEXT:    v_readlane_b32 s101, v17, 5
+; GFX11-NEXT:    v_readlane_b32 s96, v17, 0
+; GFX11-NEXT:    v_readlane_b32 s87, v16, 31
+; GFX11-NEXT:    v_readlane_b32 s85, v16, 29
+; GFX11-NEXT:    v_readlane_b32 s84, v16, 28
+; GFX11-NEXT:    v_readlane_b32 s83, v16, 27
+; GFX11-NEXT:    v_readlane_b32 s82, v16, 26
+; GFX11-NEXT:    v_readlane_b32 s81, v16, 25
+; GFX11-NEXT:    v_readlane_b32 s80, v16, 24
+; GFX11-NEXT:    v_readlane_b32 s71, v16, 23
+; GFX11-NEXT:    v_readlane_b32 s70, v16, 22
+; GFX11-NEXT:    v_readlane_b32 s68, v16, 20
+; GFX11-NEXT:    v_readlane_b32 s67, v16, 19
+; GFX11-NEXT:    v_readlane_b32 s66, v16, 18
+; GFX11-NEXT:    v_readlane_b32 s65, v16, 17
+; GFX11-NEXT:    v_readlane_b32 s64, v16, 16
+; GFX11-NEXT:    v_readlane_b32 s55, v16, 15
+; GFX11-NEXT:    v_readlane_b32 s54, v16, 14
+; GFX11-NEXT:    v_readlane_b32 s53, v16, 13
+; GFX11-NEXT:    v_readlane_b32 s52, v16, 12
+; GFX11-NEXT:    v_readlane_b32 s51, v16, 11
+; GFX11-NEXT:    v_readlane_b32 s50, v16, 10
+; GFX11-NEXT:    v_readlane_b32 s49, v16, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v16, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v16, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v16, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v16, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v16, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v16, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v16, 2
+; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_load_b32 v16, off, s32
+; GFX11-NEXT:    scratch_load_b32 v17, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v18, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v19, off, s32 offset:12
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -21647,462 +20624,482 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v32i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:436
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:320
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v86, v0
+; GFX11-TRUE16-NEXT:    s_clause 0x1f
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:476
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:472
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:468
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:464
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:460
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:456
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:452
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:448
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:444
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:352
+; GFX11-TRUE16-NEXT:    s_clause 0x7
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:320
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v26 :: v_dual_mov_b32 v35, v24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v22 :: v_dual_mov_b32 v37, v20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v18 :: v_dual_mov_b32 v39, v16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v14 :: v_dual_mov_b32 v49, v12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v10 :: v_dual_mov_b32 v51, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v6 :: v_dual_mov_b32 v53, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, v2 :: v_dual_mov_b32 v55, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
 ; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:316
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:312
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:308
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:304
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:300
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:296
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:292
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v2, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v16, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v18, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v20, off, s32 offset:72
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v22, off, s32 offset:80
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v24, off, s32 offset:88
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v26, off, s32 offset:96
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v28, off, s32 offset:104
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v30, off, s32 offset:112
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:120
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:128
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:136
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:144
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:152
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:160
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:168
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:176
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:184
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:200
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:208
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:216
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:224
 ; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v94, off, s32 offset:240
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v95, off, s32 offset:248
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v104, off, s32 offset:256
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v105, off, s32 offset:264
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v106, off, s32 offset:272
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v107, off, s32 offset:280
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v108, off, s32 offset:288
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v109, off, s32 offset:296
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v110, off, s32 offset:304
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v111, off, s32 offset:312
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:308
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:300
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:292
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:284
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:276
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:268
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:260
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:252
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:244
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:228
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:220
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:212
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:204
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:196
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:188
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:180
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:172
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:164
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:156
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:148
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:140
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:132
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v89, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v88, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v90, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v91, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v93, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v92, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v28 :: v_dual_mov_b32 v34, v27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v26 :: v_dual_mov_b32 v37, v25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v24 :: v_dual_mov_b32 v38, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v22 :: v_dual_mov_b32 v49, v21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v20 :: v_dual_mov_b32 v51, v19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v18 :: v_dual_mov_b32 v53, v17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v55, v16 :: v_dual_mov_b32 v54, v15
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, v14 :: v_dual_mov_b32 v65, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v67, v12 :: v_dual_mov_b32 v66, v11
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v68, v10 :: v_dual_mov_b32 v69, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v71, v8 :: v_dual_mov_b32 v70, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v80, v6 :: v_dual_mov_b32 v81, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v83, v4 :: v_dual_mov_b32 v82, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v84, v2 :: v_dual_mov_b32 v85, v1
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:124
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:116
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:108
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:100
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:92
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:84
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:76
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:68
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v89, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v90, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v91, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v92, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v93, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v76, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v77, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v78, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v79, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v88, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v63, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v72, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v73, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v74, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v75, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v57, 8, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v58, 8, v4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v59, 8, v6
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v60, 8, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v61, 8, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v40, 8, v12
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 8, v14
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v42, 8, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v43, 8, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v44, 8, v20
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v167, 8, v22
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v176, 8, v24
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v177, 8, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v178, 8, v28
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v179, 8, v30
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v149, 8, v31
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v150, 8, v87
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(61)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v151, 8, v96
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(60)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v160, 8, v97
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(59)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v161, 8, v98
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(58)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v132, 8, v99
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(57)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v133, 8, v113
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(56)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v134, 8, v114
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(55)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v135, 8, v115
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(54)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v144, 8, v116
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(53)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v119, 8, v117
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(52)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v128, 8, v128
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(51)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v129, 8, v129
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(50)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v130, 8, v130
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(49)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v131, 8, v131
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(48)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v113, 8, v94
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(47)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v114, 8, v95
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(46)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v115, 8, v104
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(45)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v116, 8, v105
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(44)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v117, 8, v106
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(43)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v107
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(42)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v108
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(41)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v109
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(40)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v98, 8, v110
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(39)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v99, 8, v111
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB15_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v71
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v53
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v90
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v91
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v68
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v65
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v67
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v53
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s26, 0xff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v49
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s27, 8
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v76
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v77
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v64
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v78
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v79
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v49
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v51
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v37
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v88
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v63
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v37
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v72
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v73
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v33
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v33
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v74
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v75
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v91
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v0, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v32
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v92
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v93
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v62
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v57
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v58
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v79
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v90
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v88
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v89
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v56
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v47
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v59
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v60
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v75
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v78
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v77
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v76
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v45
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v61
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v40
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v63
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v74
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v72
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v73
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v183
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v182
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v41
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v42
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v59
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v62
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v60
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v61
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v43
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v44
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v47
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v56
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v57
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v165
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v167
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v176
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v43
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v44
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v45
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v164
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v163
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v177
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v178
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v183
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v42
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v41
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v40
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v148
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v179
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v149
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v179
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v182
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v147
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v146
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v150
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v151
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v167
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v178
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v176
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v177
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v118
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v160
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v161
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v163
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v164
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v165
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v103
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v132
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v133
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v151
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v160
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v161
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v102
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v101
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v134
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v135
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v147
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v150
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v149
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v148
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v100
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v86
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v144
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v119
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v135
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v146
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v144
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v85
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v84
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v128
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v129
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v131
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v134
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v132
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v133
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v83
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v130
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v131
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v119
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v130
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v128
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v129
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v81
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v80
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v113
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v114
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v114
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v118
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v116
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v117
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v115
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v116
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v102
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v115
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v113
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v68
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v117
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v87
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v98
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v103
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v100
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v67
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v96
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v97
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v87
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v99
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v96
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v97
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v64
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v98
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v99
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v86
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v85
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v89
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
@@ -22135,8 +21132,18 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s10, 16
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v51
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v93
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v92
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB15_3
@@ -22154,445 +21161,401 @@ define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s4
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
 ; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v86
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v55
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v54
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v52
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v85
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v51
 ; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v84
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v81
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v82
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v38
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v89, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v90, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v91, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v92, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v93, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v69
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v88, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v74, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v75, v12
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v68
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v65
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v64
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v8, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v54
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v49
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v48
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v39
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v37
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v8, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v46
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v181
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v76, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v77, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v78, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v79, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v63, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v72, v8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v73, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v49
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v35
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v9, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v33
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v3, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v93
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v61, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v43, v16
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v44, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v14, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v32
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v92
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v62
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v56
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v47
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v45
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v183
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v182
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v91
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v90
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v89
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v12, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v13, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v88
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v79
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v78
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v162
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v145
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v118
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v57, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v58, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v59, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v60, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v40, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v41, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v42, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v12, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v13, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v77
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v76
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v75
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v74
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v73
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v14, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v72
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v63
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v62
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v61
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v14, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v15, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v59
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v17, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v15
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v58
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v3, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v57
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v56
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v179, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v160, v21
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v161, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v18
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v19, v15
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v166
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v165
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v164
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v163
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v148
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v147
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v146
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v47
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v46
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v45
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v17, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v18, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v44
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v43
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v42
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v100
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v83
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v167, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v176, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v177, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v178, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v149, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v150, v18
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v151, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v17, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v18, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v41
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v40
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v183
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v182
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v181
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v19, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v180
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v178
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v177
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v19, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v20, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v176
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v167
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v22, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v166
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v3, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v165
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v164
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v144, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v130, v26
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v131, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v19, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v24, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v112
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v103
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v102
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v86
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v85
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v163
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v162
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v161
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v22, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v23, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v160
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v151
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v150
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v26, v27
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v69
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v132, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v133, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v134, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v135, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v119, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v128, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v129, v25
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v22, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v23, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v149
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v148
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v147
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v146
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v145
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v24, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v144
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v135
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v134
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v133
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v24, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v25, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v132
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v131
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v27, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v130
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v3, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v128
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v129
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v119
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v113
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v27, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v115
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v117, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v24, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v29, v25
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v81
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v80
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v71
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v70
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v27
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v68
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v67
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v66
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v65
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 3, v64
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v118
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v117
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v114
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v28, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v116
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v30, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v101
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 3, v99
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v28, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v29, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v112
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v103
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v97
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v102
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xff, v30
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v98
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v87
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v28, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v33, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v35, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v36, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v37, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v113, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v114, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v115, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v116, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v87, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v96, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v97, v30
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v98, v31
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v99, v32
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v27
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v2, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v1, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v30, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v34, v35
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v29, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v34, v30
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v31, v32
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-TRUE16-NEXT:  .LBB15_3: ; %end
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:320
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    s_clause 0x1f
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:320
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:352
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:444
+; GFX11-TRUE16-NEXT:    s_clause 0x7
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:448
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:452
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:456
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:460
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:464
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:468
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:472
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:476
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ; GFX11-TRUE16-NEXT:  .LBB15_4:
@@ -30107,1735 +29070,870 @@ define inreg <32 x i32> @bitcast_v64bf16_to_v32i32_scalar(<64 x bfloat> inreg %a
 ; GFX9-NEXT:  .LBB19_4:
 ; GFX9-NEXT:    s_branch .LBB19_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v32i32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:192
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:164
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v120, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v121, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v122, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v123, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v124, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v125, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v126, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v127, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v136, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v137, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v138, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v139, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v140, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v141, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v142, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v143, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v152, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v153, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v154, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v155, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v156, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v157, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v158, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v159, s32 offset:36
-; GFX11-TRUE16-NEXT:    s_clause 0x8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v168, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v169, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v170, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v171, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v172, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v173, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v174, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v175, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v184, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v183, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v182, v9 :: v_dual_mov_b32 v169, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v171, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v176, v3 :: v_dual_mov_b32 v175, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v173, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v181, s28 :: v_dual_mov_b32 v172, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB19_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB19_3
-; GFX11-TRUE16-NEXT:  .LBB19_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s27, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s27, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s26, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s25, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s25, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff0000, v181
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v8, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v8
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v10, v6
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s24, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s23, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v7, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s22, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v11, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s22, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v9, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s21, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v14, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v11
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v16, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s20, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v12, v16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s20, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s19, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v18, v12
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v16, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s18, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v21, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v18
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s17, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v17, v20
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v19, v22
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s17, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v19, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s16, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, v24, v19
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v25, 16, 1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s16, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, v22, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v19
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, 0x400000, v25
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s3, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, v27, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v23
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v26, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v28, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s2, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, v25, v28
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v30, v25, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s1, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, v30, v25
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v31, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, v28, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v31
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v30, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v33, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, 0x400000, v29
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v30
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v29, v32, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v31, v34, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, v29, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, v31, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v31, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v37, v33
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v35, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v33, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v36, v37
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v183
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v35, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v183
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v37, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v182
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v182
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v35, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v36, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v183, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v38, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v39, v36
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v182, v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v177
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v35, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v177
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v32, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v49, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v49, v35
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v37, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v31, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v176
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v31, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v177, v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v50, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v176
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v49, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v50, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v37, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v48, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, v39, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v176, v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v34, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v37, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v39, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, v38, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v181
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v36, v49, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v36, v38
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, 0x400000, v48
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v50, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v39, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, 0x400000, v39
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v51, v48, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, v38, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v37
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v52, v50, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, v51, v48
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v52, v52, v50
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v52, 0x400000, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v181, v39, 16, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
-; GFX11-TRUE16-NEXT:  .LBB19_3: ; %end
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v181 :: v_dual_mov_b32 v19, v175
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v21, v176
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v184, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v175, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v174, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v173, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v172, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v171, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v170, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v169, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v168, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v159, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v158, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v157, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v156, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v155, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v154, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v153, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v152, off, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v143, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v142, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v141, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v140, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v139, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v138, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v137, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v136, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v127, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v126, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v125, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v124, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v123, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v122, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v121, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v120, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:192
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    s_clause 0x8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v177 :: v_dual_mov_b32 v27, v182
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v183 :: v_dual_mov_b32 v31, v178
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v30, v179
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB19_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
-; GFX11-TRUE16-NEXT:    s_branch .LBB19_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v32i32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:288
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:284
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:280
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:276
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:272
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:268
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:264
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:260
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:256
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:252
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:248
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:244
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:240
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:236
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:232
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:228
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:224
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:220
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:216
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:212
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:208
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:204
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:200
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:196
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:192
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:188
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:184
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:180
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:176
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:172
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:168
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:164
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:160
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:156
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:152
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:148
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:144
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:140
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:136
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:132
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:128
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:124
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:120
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:116
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:112
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:108
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:104
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:100
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v136, s32 offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v137, s32 offset:92
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v138, s32 offset:88
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v139, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v140, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v141, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v142, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v143, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v152, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v153, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v154, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v155, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v156, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v157, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v158, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v159, s32 offset:36
-; GFX11-FAKE16-NEXT:    s_clause 0x8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v168, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v169, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v170, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v171, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v172, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v173, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v174, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v175, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v184, s32
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v181, v9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v182, v10 :: v_dual_mov_b32 v169, v7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v176, v6 :: v_dual_mov_b32 v171, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v173, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v175, v1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v183, s28 :: v_dual_mov_b32 v172, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB19_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB19_3
-; GFX11-FAKE16-NEXT:  .LBB19_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s27, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s27, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s26, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s25, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s25, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s24, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff0000, v183
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v8, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v3, v8
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v10, v6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s24, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s23, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v7, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s23, 16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s22, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v11, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s22, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v9, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v10, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s21, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v14, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s21, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v11
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v16, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s20, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, v12, v16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s20, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s19, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v18, v12
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s19, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v16, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s18, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s18, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v18
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s17, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v17, v20
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v19, v22
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s17, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v19, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s16, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v24, v19
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v25, 16, 1
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v22, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v19
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v25
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s3, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, v27, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v23
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v26, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v25, v28, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s2, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, v25, v28
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, 0x400000, v28
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v30, v25, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s1, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, v30, v25
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v31, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, v28, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v31
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v30, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v33, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, 0x400000, v29
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, 0x400000, v30
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v29, v32, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v31, v34, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, v29, v32
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, v31, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v31, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v37, v33
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v35, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v33, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v36, v37
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v182
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v35, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v182
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v37, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v181
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v181
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v35, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v36, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v182, v31, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v38, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v39, v36
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v181, v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v176
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v35, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v176
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v32, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v49, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v37, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v31, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v177
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v31, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v176, v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v177
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v50, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v37, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v48, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, v39, v38
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v177, v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v34, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v37, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v39, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, v38, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v36, v49 :: v_dual_lshlrev_b32 v48, 16, v183
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v36, v38
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v48
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v39, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v39
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v48, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, v38, v39
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v37
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v50, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v51, v51, v48
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v52, v52, v50
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, 0x400000, v50
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v183, v39, 16, v48
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
-; GFX11-FAKE16-NEXT:  .LBB19_3: ; %end
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v19, v175
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v184, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v175, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v174, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v173, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v172, off, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v171, off, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v170, off, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v169, off, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v168, off, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v159, off, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v158, off, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v157, off, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v156, off, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v155, off, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v154, off, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v153, off, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v152, off, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v143, off, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v142, off, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v141, off, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v140, off, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v139, off, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v138, off, s32 offset:88
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v137, off, s32 offset:92
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v136, off, s32 offset:96
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:100
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:104
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:108
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:112
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:116
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:120
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:124
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:128
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:132
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:136
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:140
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:144
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:148
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:152
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:156
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:160
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:164
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:168
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:172
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:176
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:180
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:184
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:188
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:192
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:196
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:200
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:204
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:208
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:212
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:216
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:220
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:224
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:228
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:232
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:236
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:240
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:244
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:248
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:252
-; GFX11-FAKE16-NEXT:    s_clause 0x8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:256
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:260
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:264
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:268
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:272
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:276
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:280
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:284
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:288
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v183 :: v_dual_mov_b32 v21, v177
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, v176 :: v_dual_mov_b32 v27, v181
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v28, v182
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, v179 :: v_dual_mov_b32 v31, v178
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB19_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
-; GFX11-FAKE16-NEXT:    s_branch .LBB19_2
+; GFX11-LABEL: bitcast_v64bf16_to_v32i32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:288
+; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:284
+; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:280
+; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:276
+; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:272
+; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:268
+; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:264
+; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:260
+; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:256
+; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:252
+; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:248
+; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:244
+; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:240
+; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:236
+; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:232
+; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:228
+; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:224
+; GFX11-NEXT:    scratch_store_b32 off, v73, s32 offset:220
+; GFX11-NEXT:    scratch_store_b32 off, v74, s32 offset:216
+; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:212
+; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:208
+; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:204
+; GFX11-NEXT:    scratch_store_b32 off, v78, s32 offset:200
+; GFX11-NEXT:    scratch_store_b32 off, v79, s32 offset:196
+; GFX11-NEXT:    scratch_store_b32 off, v88, s32 offset:192
+; GFX11-NEXT:    scratch_store_b32 off, v89, s32 offset:188
+; GFX11-NEXT:    scratch_store_b32 off, v90, s32 offset:184
+; GFX11-NEXT:    scratch_store_b32 off, v91, s32 offset:180
+; GFX11-NEXT:    scratch_store_b32 off, v92, s32 offset:176
+; GFX11-NEXT:    scratch_store_b32 off, v93, s32 offset:172
+; GFX11-NEXT:    scratch_store_b32 off, v94, s32 offset:168
+; GFX11-NEXT:    scratch_store_b32 off, v95, s32 offset:164
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_store_b32 off, v104, s32 offset:160
+; GFX11-NEXT:    scratch_store_b32 off, v105, s32 offset:156
+; GFX11-NEXT:    scratch_store_b32 off, v106, s32 offset:152
+; GFX11-NEXT:    scratch_store_b32 off, v107, s32 offset:148
+; GFX11-NEXT:    scratch_store_b32 off, v108, s32 offset:144
+; GFX11-NEXT:    scratch_store_b32 off, v109, s32 offset:140
+; GFX11-NEXT:    scratch_store_b32 off, v110, s32 offset:136
+; GFX11-NEXT:    scratch_store_b32 off, v111, s32 offset:132
+; GFX11-NEXT:    scratch_store_b32 off, v120, s32 offset:128
+; GFX11-NEXT:    scratch_store_b32 off, v121, s32 offset:124
+; GFX11-NEXT:    scratch_store_b32 off, v122, s32 offset:120
+; GFX11-NEXT:    scratch_store_b32 off, v123, s32 offset:116
+; GFX11-NEXT:    scratch_store_b32 off, v124, s32 offset:112
+; GFX11-NEXT:    scratch_store_b32 off, v125, s32 offset:108
+; GFX11-NEXT:    scratch_store_b32 off, v126, s32 offset:104
+; GFX11-NEXT:    scratch_store_b32 off, v127, s32 offset:100
+; GFX11-NEXT:    scratch_store_b32 off, v136, s32 offset:96
+; GFX11-NEXT:    scratch_store_b32 off, v137, s32 offset:92
+; GFX11-NEXT:    scratch_store_b32 off, v138, s32 offset:88
+; GFX11-NEXT:    scratch_store_b32 off, v139, s32 offset:84
+; GFX11-NEXT:    scratch_store_b32 off, v140, s32 offset:80
+; GFX11-NEXT:    scratch_store_b32 off, v141, s32 offset:76
+; GFX11-NEXT:    scratch_store_b32 off, v142, s32 offset:72
+; GFX11-NEXT:    scratch_store_b32 off, v143, s32 offset:68
+; GFX11-NEXT:    scratch_store_b32 off, v152, s32 offset:64
+; GFX11-NEXT:    scratch_store_b32 off, v153, s32 offset:60
+; GFX11-NEXT:    scratch_store_b32 off, v154, s32 offset:56
+; GFX11-NEXT:    scratch_store_b32 off, v155, s32 offset:52
+; GFX11-NEXT:    scratch_store_b32 off, v156, s32 offset:48
+; GFX11-NEXT:    scratch_store_b32 off, v157, s32 offset:44
+; GFX11-NEXT:    scratch_store_b32 off, v158, s32 offset:40
+; GFX11-NEXT:    scratch_store_b32 off, v159, s32 offset:36
+; GFX11-NEXT:    s_clause 0x8
+; GFX11-NEXT:    scratch_store_b32 off, v168, s32 offset:32
+; GFX11-NEXT:    scratch_store_b32 off, v169, s32 offset:28
+; GFX11-NEXT:    scratch_store_b32 off, v170, s32 offset:24
+; GFX11-NEXT:    scratch_store_b32 off, v171, s32 offset:20
+; GFX11-NEXT:    scratch_store_b32 off, v172, s32 offset:16
+; GFX11-NEXT:    scratch_store_b32 off, v173, s32 offset:12
+; GFX11-NEXT:    scratch_store_b32 off, v174, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v175, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v184, s32
+; GFX11-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
+; GFX11-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v181, v9
+; GFX11-NEXT:    v_dual_mov_b32 v182, v10 :: v_dual_mov_b32 v169, v7
+; GFX11-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v3
+; GFX11-NEXT:    v_dual_mov_b32 v176, v6 :: v_dual_mov_b32 v171, v4
+; GFX11-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v173, v0
+; GFX11-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v175, v1
+; GFX11-NEXT:    v_dual_mov_b32 v183, s28 :: v_dual_mov_b32 v172, s29
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB19_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
+; GFX11-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
+; GFX11-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
+; GFX11-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
+; GFX11-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
+; GFX11-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
+; GFX11-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
+; GFX11-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB19_3
+; GFX11-NEXT:  .LBB19_2: ; %cmp.true
+; GFX11-NEXT:    s_and_b32 s5, s27, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s4, s27, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s6, s26, 16
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v1
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_bfe_u32 v10, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_bfe_u32 v9, v3, 16, 1
+; GFX11-NEXT:    s_lshl_b32 s7, s25, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    s_and_b32 s5, s25, 0xffff0000
+; GFX11-NEXT:    s_and_b32 s4, s24, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
+; GFX11-NEXT:    v_and_b32_e32 v51, 0xffff0000, v183
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v10, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v9, v3
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_bfe_u32 v10, v6, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_bfe_u32 v3, v8, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v3, v8
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v10, v6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    s_lshl_b32 s4, s24, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v6
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v8
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
+; GFX11-NEXT:    s_and_b32 s4, s23, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v4, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v5
+; GFX11-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
+; GFX11-NEXT:    v_bfe_u32 v7, v9, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, v7, v9
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s23, 16
+; GFX11-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_bfe_u32 v11, v7, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    s_and_b32 s4, s22, 0xffff0000
+; GFX11-NEXT:    v_bfe_u32 v9, v12, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v11, v7
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s22, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
+; GFX11-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v9, v12
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v7
+; GFX11-NEXT:    v_bfe_u32 v14, v10, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_or_b32_e32 v13, 0x400000, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
+; GFX11-NEXT:    s_and_b32 s4, s21, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v9, v11, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, v14, v10
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
+; GFX11-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v10
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    s_lshl_b32 s4, s21, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v11
+; GFX11-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v10, v13, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_bfe_u32 v12, v16, 16, 1
+; GFX11-NEXT:    s_and_b32 s4, s20, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, v12, v16
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    s_lshl_b32 s4, s20, 16
+; GFX11-NEXT:    v_or_b32_e32 v17, 0x400000, v16
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
+; GFX11-NEXT:    v_bfe_u32 v18, v12, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
+; GFX11-NEXT:    s_and_b32 s4, s19, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v18, v12
+; GFX11-NEXT:    v_bfe_u32 v16, v19, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v9
+; GFX11-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s19, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v16, v19
+; GFX11-NEXT:    v_or_b32_e32 v16, 0x400000, v12
+; GFX11-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v21, v17, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
+; GFX11-NEXT:    v_or_b32_e32 v20, 0x400000, v19
+; GFX11-NEXT:    s_and_b32 s4, s18, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v16, v18, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
+; GFX11-NEXT:    v_add_nc_u32_e32 v19, v21, v17
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
+; GFX11-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
+; GFX11-NEXT:    v_or_b32_e32 v19, 0x400000, v17
+; GFX11-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
+; GFX11-NEXT:    s_lshl_b32 s4, s18, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
+; GFX11-NEXT:    v_or_b32_e32 v21, 0x400000, v18
+; GFX11-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v17, v20, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_bfe_u32 v19, v22, 16, 1
+; GFX11-NEXT:    s_and_b32 s4, s17, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v17, v17, v20
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
+; GFX11-NEXT:    v_or_b32_e32 v21, 0x400000, v20
+; GFX11-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
+; GFX11-NEXT:    v_add_nc_u32_e32 v18, v19, v22
+; GFX11-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
+; GFX11-NEXT:    s_lshl_b32 s4, s17, 16
+; GFX11-NEXT:    v_or_b32_e32 v23, 0x400000, v22
+; GFX11-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
+; GFX11-NEXT:    v_bfe_u32 v24, v19, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
+; GFX11-NEXT:    s_and_b32 s4, s16, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, v24, v19
+; GFX11-NEXT:    v_bfe_u32 v22, v25, 16, 1
+; GFX11-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
+; GFX11-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s16, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
+; GFX11-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, v22, v25
+; GFX11-NEXT:    v_or_b32_e32 v22, 0x400000, v19
+; GFX11-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v27, v23, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
+; GFX11-NEXT:    v_or_b32_e32 v26, 0x400000, v25
+; GFX11-NEXT:    s_and_b32 s4, s3, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v22, v24, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
+; GFX11-NEXT:    v_add_nc_u32_e32 v25, v27, v23
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
+; GFX11-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
+; GFX11-NEXT:    v_or_b32_e32 v25, 0x400000, v23
+; GFX11-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
+; GFX11-NEXT:    v_or_b32_e32 v27, 0x400000, v24
+; GFX11-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
+; GFX11-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v23, v26, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
+; GFX11-NEXT:    v_bfe_u32 v25, v28, 16, 1
+; GFX11-NEXT:    s_and_b32 s3, s2, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
+; GFX11-NEXT:    v_or_b32_e32 v27, 0x400000, v26
+; GFX11-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v24, v25, v28
+; GFX11-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_or_b32_e32 v29, 0x400000, v28
+; GFX11-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
+; GFX11-NEXT:    v_bfe_u32 v30, v25, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
+; GFX11-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
+; GFX11-NEXT:    s_and_b32 s2, s1, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, v30, v25
+; GFX11-NEXT:    v_bfe_u32 v28, v31, 16, 1
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, v28, v31
+; GFX11-NEXT:    v_or_b32_e32 v28, 0x400000, v25
+; GFX11-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
+; GFX11-NEXT:    v_bfe_u32 v33, v29, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
+; GFX11-NEXT:    v_or_b32_e32 v32, 0x400000, v31
+; GFX11-NEXT:    s_and_b32 s1, s0, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v28, v30, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v33, v29
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
+; GFX11-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
+; GFX11-NEXT:    v_or_b32_e32 v31, 0x400000, v29
+; GFX11-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
+; GFX11-NEXT:    v_or_b32_e32 v33, 0x400000, v30
+; GFX11-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
+; GFX11-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v29, v32, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
+; GFX11-NEXT:    v_or_b32_e32 v35, 0x400000, v32
+; GFX11-NEXT:    v_bfe_u32 v31, v34, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
+; GFX11-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, v29, v32
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
+; GFX11-NEXT:    v_add_nc_u32_e32 v30, v31, v34
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
+; GFX11-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
+; GFX11-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
+; GFX11-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v37, v33, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v29
+; GFX11-NEXT:    v_bfe_u32 v35, v31, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v37, v33
+; GFX11-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v35, v31
+; GFX11-NEXT:    v_or_b32_e32 v37, 0x400000, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_or_b32_e32 v38, 0x400000, v31
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v34, v36, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v33, v35, 16, 1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v33, v35
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_bfe_u32 v36, v37, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_bfe_u32 v35, v38, 16, 1
+; GFX11-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v36, v37
+; GFX11-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v182
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v35, v38
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v182
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xffff, v30
+; GFX11-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_bfe_u32 v37, v36, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v37, v36
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v181
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v181
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-NEXT:    v_bfe_u32 v35, v37, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v36, v38, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v35, v37
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v36, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v37
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v39, v36, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_lshl_or_b32 v182, v31, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v38, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v39, v36
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v38, 0x400000, v36
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
+; GFX11-NEXT:    v_lshl_or_b32 v181, v32, 16, v33
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff0000, v176
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_bfe_u32 v35, v37, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
+; GFX11-NEXT:    v_bfe_u32 v32, v36, 16, 1
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v35, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 16, v176
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v32, v36
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_bfe_u32 v37, v38, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v49, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v49, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v37, v35, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
+; GFX11-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v37, v35
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v31, v36, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v177
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v31, v36
+; GFX11-NEXT:    v_lshl_or_b32 v176, v33, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
+; GFX11-NEXT:    v_or_b32_e32 v34, 0x400000, v35
+; GFX11-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v34, 0x400000, v36
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
+; GFX11-NEXT:    v_bfe_u32 v37, v32, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v32
+; GFX11-NEXT:    v_bfe_u32 v50, v38, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
+; GFX11-NEXT:    v_bfe_u32 v34, v35, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v177
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_bfe_u32 v49, v37, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v50, v38
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v37
+; GFX11-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_bfe_u32 v48, v37, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v39, v38, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
+; GFX11-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v48, v37
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
+; GFX11-NEXT:    v_add_nc_u32_e32 v39, v39, v38
+; GFX11-NEXT:    v_lshl_or_b32 v177, v35, 16, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_or_b32_e32 v35, 0x400000, v37
+; GFX11-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v37, v33, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v39, v34, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v37, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v39, v34
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, v38, v35
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v33, v36, v49 :: v_dual_lshlrev_b32 v48, 16, v183
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v36, v38, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v38
+; GFX11-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v36, v38
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_or_b32_e32 v55, 0x400000, v48
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_bfe_u32 v50, v37, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v38, v39, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v54, 0x400000, v39
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
+; GFX11-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
+; GFX11-NEXT:    v_bfe_u32 v51, v48, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, v38, v39
+; GFX11-NEXT:    v_or_b32_e32 v53, 0x400000, v37
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
+; GFX11-NEXT:    v_bfe_u32 v52, v50, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v51, v51, v48
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v52, v52, v50
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
+; GFX11-NEXT:    v_or_b32_e32 v52, 0x400000, v50
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
+; GFX11-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
+; GFX11-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
+; GFX11-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
+; GFX11-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
+; GFX11-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
+; GFX11-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
+; GFX11-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
+; GFX11-NEXT:    v_lshl_or_b32 v183, v39, 16, v48
+; GFX11-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
+; GFX11-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
+; GFX11-NEXT:  .LBB19_3: ; %end
+; GFX11-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
+; GFX11-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
+; GFX11-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
+; GFX11-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
+; GFX11-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
+; GFX11-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
+; GFX11-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v19, v175
+; GFX11-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
+; GFX11-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
+; GFX11-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_load_b32 v184, off, s32
+; GFX11-NEXT:    scratch_load_b32 v175, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v174, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v173, off, s32 offset:12
+; GFX11-NEXT:    scratch_load_b32 v172, off, s32 offset:16
+; GFX11-NEXT:    scratch_load_b32 v171, off, s32 offset:20
+; GFX11-NEXT:    scratch_load_b32 v170, off, s32 offset:24
+; GFX11-NEXT:    scratch_load_b32 v169, off, s32 offset:28
+; GFX11-NEXT:    scratch_load_b32 v168, off, s32 offset:32
+; GFX11-NEXT:    scratch_load_b32 v159, off, s32 offset:36
+; GFX11-NEXT:    scratch_load_b32 v158, off, s32 offset:40
+; GFX11-NEXT:    scratch_load_b32 v157, off, s32 offset:44
+; GFX11-NEXT:    scratch_load_b32 v156, off, s32 offset:48
+; GFX11-NEXT:    scratch_load_b32 v155, off, s32 offset:52
+; GFX11-NEXT:    scratch_load_b32 v154, off, s32 offset:56
+; GFX11-NEXT:    scratch_load_b32 v153, off, s32 offset:60
+; GFX11-NEXT:    scratch_load_b32 v152, off, s32 offset:64
+; GFX11-NEXT:    scratch_load_b32 v143, off, s32 offset:68
+; GFX11-NEXT:    scratch_load_b32 v142, off, s32 offset:72
+; GFX11-NEXT:    scratch_load_b32 v141, off, s32 offset:76
+; GFX11-NEXT:    scratch_load_b32 v140, off, s32 offset:80
+; GFX11-NEXT:    scratch_load_b32 v139, off, s32 offset:84
+; GFX11-NEXT:    scratch_load_b32 v138, off, s32 offset:88
+; GFX11-NEXT:    scratch_load_b32 v137, off, s32 offset:92
+; GFX11-NEXT:    scratch_load_b32 v136, off, s32 offset:96
+; GFX11-NEXT:    scratch_load_b32 v127, off, s32 offset:100
+; GFX11-NEXT:    scratch_load_b32 v126, off, s32 offset:104
+; GFX11-NEXT:    scratch_load_b32 v125, off, s32 offset:108
+; GFX11-NEXT:    scratch_load_b32 v124, off, s32 offset:112
+; GFX11-NEXT:    scratch_load_b32 v123, off, s32 offset:116
+; GFX11-NEXT:    scratch_load_b32 v122, off, s32 offset:120
+; GFX11-NEXT:    scratch_load_b32 v121, off, s32 offset:124
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_load_b32 v120, off, s32 offset:128
+; GFX11-NEXT:    scratch_load_b32 v111, off, s32 offset:132
+; GFX11-NEXT:    scratch_load_b32 v110, off, s32 offset:136
+; GFX11-NEXT:    scratch_load_b32 v109, off, s32 offset:140
+; GFX11-NEXT:    scratch_load_b32 v108, off, s32 offset:144
+; GFX11-NEXT:    scratch_load_b32 v107, off, s32 offset:148
+; GFX11-NEXT:    scratch_load_b32 v106, off, s32 offset:152
+; GFX11-NEXT:    scratch_load_b32 v105, off, s32 offset:156
+; GFX11-NEXT:    scratch_load_b32 v104, off, s32 offset:160
+; GFX11-NEXT:    scratch_load_b32 v95, off, s32 offset:164
+; GFX11-NEXT:    scratch_load_b32 v94, off, s32 offset:168
+; GFX11-NEXT:    scratch_load_b32 v93, off, s32 offset:172
+; GFX11-NEXT:    scratch_load_b32 v92, off, s32 offset:176
+; GFX11-NEXT:    scratch_load_b32 v91, off, s32 offset:180
+; GFX11-NEXT:    scratch_load_b32 v90, off, s32 offset:184
+; GFX11-NEXT:    scratch_load_b32 v89, off, s32 offset:188
+; GFX11-NEXT:    scratch_load_b32 v88, off, s32 offset:192
+; GFX11-NEXT:    scratch_load_b32 v79, off, s32 offset:196
+; GFX11-NEXT:    scratch_load_b32 v78, off, s32 offset:200
+; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:204
+; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:208
+; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:212
+; GFX11-NEXT:    scratch_load_b32 v74, off, s32 offset:216
+; GFX11-NEXT:    scratch_load_b32 v73, off, s32 offset:220
+; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:224
+; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:228
+; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:232
+; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:236
+; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:240
+; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:244
+; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:248
+; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:252
+; GFX11-NEXT:    s_clause 0x8
+; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:256
+; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:260
+; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:264
+; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:268
+; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:272
+; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:276
+; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:280
+; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:284
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:288
+; GFX11-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
+; GFX11-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
+; GFX11-NEXT:    v_dual_mov_b32 v16, v183 :: v_dual_mov_b32 v21, v177
+; GFX11-NEXT:    v_dual_mov_b32 v24, v176 :: v_dual_mov_b32 v27, v181
+; GFX11-NEXT:    v_mov_b32_e32 v28, v182
+; GFX11-NEXT:    v_dual_mov_b32 v30, v179 :: v_dual_mov_b32 v31, v178
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB19_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
+; GFX11-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
+; GFX11-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
+; GFX11-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
+; GFX11-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
+; GFX11-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
+; GFX11-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
+; GFX11-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
+; GFX11-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
+; GFX11-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
+; GFX11-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
+; GFX11-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
+; GFX11-NEXT:    s_branch .LBB19_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -49034,2095 +47132,1035 @@ define inreg <128 x i8> @bitcast_v32f32_to_v128i8_scalar(<32 x float> inreg %a,
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32f32_to_v128i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:84
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s30, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s96, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s97, 1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s98, 2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s99, 3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s100, 4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s101, 5
-; GFX11-TRUE16-NEXT:    s_mov_b32 vcc_hi, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_clause 0x11
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s102, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s103, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s104, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s50, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s51, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s52, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s53, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s54, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s55, 15
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s64, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s65, 17
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s66, 18
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s67, 19
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s68, 20
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s69, 21
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s70, 22
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s71, 23
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s80, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s81, 25
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s82, 26
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s83, 27
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s84, 28
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s85, 29
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s86, 30
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s87, 31
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB37_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s26, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s25, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s51, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s52, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s53, s23, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s12, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s54, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s55, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s64, s22, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s65, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s66, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s67, s21, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s68, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s69, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s70, s20, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s71, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s80, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s81, s19, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s82, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s83, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s84, s17, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s85, s17, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 31
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s86, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s87, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s96, s16, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 30
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s97, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s98, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s3, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 29
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s100, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s101, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s1, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 28
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s0, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 27
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[72:73], s[10:11], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 26
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[8:9], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[78:79], s[4:5], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 25
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[88:89], s[14:15], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[90:91], s[40:41], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[28:29], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 23
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 22
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s6, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 21
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 20
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 19
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 18
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s4, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 17
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s4, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 15
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 14
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 13
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s14, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s14, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 11
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s40, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s40, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s29, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s29, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s29, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB37_4
-; GFX11-TRUE16-NEXT:  .LBB37_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v31, s25, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v30, s24, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, s23, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, s22, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, s21, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, s20, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v37, s19, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v36, s18, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v39, s17, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v38, s16, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, s5, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, s4, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v51, s3, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v50, s2, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, s15, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, s14, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, s9, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, s8, 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[30:31]
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v55, s1, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v54, s0, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v29, s27, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v28, s26, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, s29, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, s28, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, s41, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, s40, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, s7, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, s11, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, s13, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, s12, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, s10, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, s6, 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[36:37]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[32:33], 24, v[13:14]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[38:39]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[25:26], 24, v[9:10]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[69:70], 24, v[50:51]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[21:22]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[23:24]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[28:29]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[54:55]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 24, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 8, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 8, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 24, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 24, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 8, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 24, v29
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v29
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 24, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 8, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v164, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v165, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v177, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v179, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v182, 24, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v183, 16, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v40, 8, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v43, 24, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v44, 16, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v45, 8, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v56, 24, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v58, 8, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v61, 24, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v62, 16, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v63, 8, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v54
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v54
-; GFX11-TRUE16-NEXT:    s_branch .LBB37_5
-; GFX11-TRUE16-NEXT:  .LBB37_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr104_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr103_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr102_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr101_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr99_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr98_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr97_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr96_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr87_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr86_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr85_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr84_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 2
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 8
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 9
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 10
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 11
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 12
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 13
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 14
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 15
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 17
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 18
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 19
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 20
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 21
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 22
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 23
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 24
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 25
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 26
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 27
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 28
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 29
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 30
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 2
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 8
-; GFX11-TRUE16-NEXT:    s_branch .LBB37_2
-; GFX11-TRUE16-NEXT:  .LBB37_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v146.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v145.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v144.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v135.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v134.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v133.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v132.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v73.l, s35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v131.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v72.l, s34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v63.l, s104
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, s103
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v61.l, s102
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v130.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, s101
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v59.l, s100
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, s99
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, s98
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v129.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v56.l, s97
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, s96
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v46.l, s87
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, s86
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v44.l, s85
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v43.l, s84
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, s83
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v41.l, s82
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v119.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, s81
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v183.l, s80
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v182.l, s71
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v181.l, s70
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v118.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v180.l, s69
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v179.l, s68
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v178.l, s67
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v177.l, s66
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v117.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v176.l, s65
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v167.l, s64
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v166.l, s55
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v165.l, s54
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v116.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v164.l, s53
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v163.l, s52
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v162.l, s51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v161.l, s50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v115.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v160.l, s49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v151.l, s48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v150.l, s39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v149.l, s38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v114.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v148.l, s37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v147.l, s36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s72
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v113.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s74
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s76
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s78
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s88
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v112.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s90
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s92
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s94
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, s30
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v103.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v68.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v102.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v69.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v70.l, s42
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v101.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v100.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 23
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v99.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v98.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 25
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v97.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v96.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 27
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v86.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 29
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 30
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 31
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v83.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v82.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v81.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v80.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s0
-; GFX11-TRUE16-NEXT:  .LBB37_5: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v73
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xff, v55
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v61, 8, v61
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, v54, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xff, v72
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xff, v50
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v60, 8, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v59, 0xff, v59
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v70, v71, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v63
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v50, v60
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, v59, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 16, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, v55, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xff, v62
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v57, 0xff, v57
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v56, 8, v56
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, v54, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v71, v71, v61
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v46, 0xff, v46
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 16, v71
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, v55, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v51, v71
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v70, v57, v56
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v47
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v56, v50, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v68, v46, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v38, v71
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v45
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xff, v44
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v43
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v57, v50, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, v39, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v70, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v42
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xff, v41
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v38, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v68, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v70, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v40
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v183
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v182
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, v38, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v67, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v69, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xff, v180
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v181
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v68, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v179
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v38, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, v50, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xff, v178
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v177
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v176
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v166
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v51, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xff, v167
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v65
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[54:57], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[36:39], off offset:16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v150
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v65, v68, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xff, v30
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v163
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xff, v162
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 8, v161
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xff, v160
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v151
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v28, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v30, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v37, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v31, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, v50, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v149
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 8, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v148
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xff, v147
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v146
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v145
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xff, v144
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v28, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v29, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, v51, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, v55, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v30, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v31, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v38, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v51, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v23, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v135
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xff, v134
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v133
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 8, v132
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xff, v131
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v130
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v37, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, v48, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v22, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xff, v129
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v128
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 8, v119
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xff, v118
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v117
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xff, v116
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v115
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, v37, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v17, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, v48, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v51, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v21, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v101
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, v23, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, v22, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v17, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v18, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v114
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v113
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v112
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v103
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v102
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xff, v87
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v18, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v32, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v100
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v98
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v97
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v48, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v99
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v32, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v14, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v9, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v86
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v84
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xff, v165
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v164
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v83
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v14, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v25, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v70, v71
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v53
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v49
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v25, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v20, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v67, v65
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v68, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v19, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v20, v16
-; GFX11-TRUE16-NEXT:    s_clause 0x5
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[28:31], off offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[36:39], off offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[11:14], off offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:112
-; GFX11-TRUE16-NEXT:    s_clause 0x11
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s104, v75, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s103, v75, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s102, v75, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s101, v75, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s100, v75, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s99, v75, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s98, v75, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s97, v75, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s96, v75, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s87, v74, 31
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s86, v74, 30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s85, v74, 29
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s84, v74, 28
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s83, v74, 27
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s82, v74, 26
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s81, v74, 25
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s80, v74, 24
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s71, v74, 23
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s70, v74, 22
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s69, v74, 21
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s68, v74, 20
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v74, 19
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v74, 18
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v74, 17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v74, 16
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v74, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v74, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v74, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v74, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v74, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v74, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v74, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v74, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v74, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v74, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v74, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v74, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v74, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v74, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v74, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v74, 0
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v32f32_to_v128i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:88
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s96, 0
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s97, 1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s98, 2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s99, 3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s100, 4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s101, 5
-; GFX11-FAKE16-NEXT:    s_mov_b32 vcc_hi, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_clause 0x12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s102, 6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s103, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s104, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s49, 9
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s50, 10
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s51, 11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s52, 12
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s53, 13
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s54, 14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s55, 15
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s64, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s65, 17
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s66, 18
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s67, 19
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s68, 20
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s69, 21
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s70, 22
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s71, 23
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s80, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s81, 25
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s82, 26
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s83, 27
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s84, 28
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s85, 29
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s86, 30
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s87, 31
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB37_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s27, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s26, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s51, s25, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s50, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s52, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s53, s23, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s12, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s54, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s64, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s55, s22, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s11, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s65, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s66, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s67, s21, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s69, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s68, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s70, s20, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s11, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s71, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s80, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s82, s19, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s81, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s83, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s84, s17, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s10, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s85, s17, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 31
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s87, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s86, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s96, s16, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 30
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s97, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s98, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s3, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 29
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s99, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s101, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s1, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 28
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s0, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 27
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s8, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[72:73], s[10:11], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 26
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s7, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[74:75], s[8:9], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[78:79], s[4:5], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 25
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[88:89], s[14:15], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[90:91], s[40:41], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[92:93], s[28:29], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s7, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 23
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 22
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s6, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 21
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 20
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 19
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 18
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s4, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 17
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s4, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s15, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 15
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s15, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 14
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s15, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 13
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s14, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s14, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 11
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s41, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 10
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s41, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s41, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s40, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s40, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s29, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s29, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s29, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB37_4
-; GFX11-FAKE16-NEXT:  .LBB37_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, s27, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, s26, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, s25, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, s24, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v29, s23, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v28, s22, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v31, s21, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v30, s20, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v35, s19, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, s18, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v37, s17, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v36, s16, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, s9, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, s8, 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[21:22]
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v53, s1, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v52, s0, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v49, s3, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v48, s2, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, s29, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, s28, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, s41, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, s40, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, s15, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, s14, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, s5, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, s7, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, s11, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, s13, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, s12, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, s10, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, s6, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, s4, 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[23:24]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[28:29]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[30:31]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[34:35]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[69:70], 24, v[36:37]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[7:8]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[32:33], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[11:12]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[48:49]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[80:81], 24, v[52:53]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 24, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 8, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 8, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 24, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 8, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 8, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 24, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 8, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v30
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v30
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 24, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 16, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 8, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 24, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 16, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 8, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 24, v49
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 16, v49
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v49
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 16, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 8, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 24, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 16, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 8, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v52
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v52
-; GFX11-FAKE16-NEXT:    s_branch .LBB37_5
-; GFX11-FAKE16-NEXT:  .LBB37_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr104
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr103
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr102
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr101
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr99
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr100
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr98
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr97
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr96
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr86
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr87
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr85
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    s_branch .LBB37_2
-; GFX11-FAKE16-NEXT:  .LBB37_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s0 :: v_dual_mov_b32 v53, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v147, s36 :: v_dual_mov_b32 v48, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v49, s3 :: v_dual_mov_b32 v36, s16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v37, s17 :: v_dual_mov_b32 v148, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s18 :: v_dual_mov_b32 v35, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s20 :: v_dual_mov_b32 v31, s21
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v146, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s22 :: v_dual_mov_b32 v29, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s24 :: v_dual_mov_b32 v24, s25
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v145, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v21, s26 :: v_dual_mov_b32 v22, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v18, s29
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v144, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v134, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s4 :: v_dual_mov_b32 v10, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s6 :: v_dual_mov_b32 v8, s7
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v135, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s10 :: v_dual_mov_b32 v4, s11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v133, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s12 :: v_dual_mov_b32 v2, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v74, s35 :: v_dual_mov_b32 v73, s104
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v132, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v72, s34 :: v_dual_mov_b32 v63, s103
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v62, s102 :: v_dual_mov_b32 v61, s101
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v131, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v60, s99 :: v_dual_mov_b32 v59, s100
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v58, s98 :: v_dual_mov_b32 v57, s97
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v129, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v56, s96 :: v_dual_mov_b32 v47, s86
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v46, s87 :: v_dual_mov_b32 v45, s85
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v130, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v44, s84 :: v_dual_mov_b32 v43, s83
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v42, s81 :: v_dual_mov_b32 v41, s82
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v128, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v40, s80 :: v_dual_mov_b32 v183, s71
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v182, s70 :: v_dual_mov_b32 v181, s68
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v119, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v180, s69 :: v_dual_mov_b32 v179, s67
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v178, s66 :: v_dual_mov_b32 v177, s65
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v118, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v176, s55 :: v_dual_mov_b32 v167, s64
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v166, s54 :: v_dual_mov_b32 v165, s53
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v116, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v164, s52 :: v_dual_mov_b32 v163, s50
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v162, s51 :: v_dual_mov_b32 v161, s49
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v117, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v160, s48 :: v_dual_mov_b32 v151, s39
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v150, s37 :: v_dual_mov_b32 v149, s38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v115, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s62 :: v_dual_mov_b32 v38, s88
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s72 :: v_dual_mov_b32 v50, s90
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v114, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, s74 :: v_dual_mov_b32 v54, s92
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v64, s94 :: v_dual_mov_b32 v65, s30
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v113, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, s60 :: v_dual_mov_b32 v67, s58
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v68, s56 :: v_dual_mov_b32 v69, s46
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v103, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 20
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v70, s44
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v80, s42
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v26, s76
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v32, s78
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v112, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 21
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v102, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 22
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v101, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 23
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v100, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 24
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v98, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 25
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v99, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 26
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v97, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 27
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v96, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 28
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v87, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 29
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v85, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 30
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v86, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 31
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v84, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v83, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v82, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v51, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v55, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v39, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v33, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v27, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v20, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v16, s0
-; GFX11-FAKE16-NEXT:  .LBB37_5: ; %end
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v80
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v74
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xff, v52
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xff, v53
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v62
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xff, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v52, v80
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v80, 0xff, v73
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v60, 0xff, v60
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v49, 0xff, v49
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xff, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v71, v80, v71
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v72
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v70, v60, v70
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v69
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v71, 16, v71
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v53, v80
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v80, 0xff, v63
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v60, v52, v71
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v53
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v80, v80, v81
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v61
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v59
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v67
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v80
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, v48, v81
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v80, 0xff, v58
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v57
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, v49, v71
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v61, v52, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v70
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v80, v81
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v56
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v71, 0xff, v47
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v62, v48, v52
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v49
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v53
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v36, v70
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v71, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v46
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xff, v45
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v44
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v63, v48, v49
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v37, v53
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, v69, v70
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xff, v34
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v43
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xff, v42
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v36, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v49
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, v52, v53
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, v69, v68
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xff, v35
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v41
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v68, 0xff, v40
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v183
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v36, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v48
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v49
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, v52, v53
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, v68, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v182
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xff, v181
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v36, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v30, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v53, v67
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v177
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v67, 0xff, v176
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v48, v49
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v49, 0xff, v179
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v52
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v178
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v180
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, v49, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v67, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v31, v68
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v167
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xff, v166
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v165
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v29, v68
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v69, v70
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v30, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v66, v52
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v151
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v67, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[60:63], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[34:37], off offset:16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v49
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v164
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xff, v163
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v65
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v162
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xff, v161
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 8, v160
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v67, v53
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v23, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v35, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v37
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v48, v49
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v150
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v64
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v149
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v146
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xff, v145
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v54
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v49, 0xff, v147
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v148
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v37
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v22, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v53
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, v64, v54
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v49, v52
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v23, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v24, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v36, v49
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v17, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v144
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xff, v134
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v135
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xff, v132
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 8, v50
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v52, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v133
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v35, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v48, v49
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v131
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v130
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xff, v129
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v128
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v49, 0xff, v119
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v118
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xff, v116
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v117
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v35, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v49, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v50
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, v52, v53
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v17, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v13, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v14, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v115
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v114
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v113
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v102
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v49
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v103
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v112
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v101
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v97
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v49, 0xff, v96
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v32, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v100
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v98
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v99
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v49, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v32, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v9, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v10, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v5, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v87
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v85
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v86
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v83
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v84
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v82
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v10, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v25, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v26
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v55
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v51
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v39
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v33
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v25, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v19, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v20, v16
-; GFX11-FAKE16-NEXT:    s_clause 0x5
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[28:31], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[34:37], off offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
-; GFX11-FAKE16-NEXT:    s_clause 0x12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:72
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s104, v76, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s103, v76, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s102, v76, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s101, v76, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s100, v76, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s99, v76, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s98, v76, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s97, v76, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s96, v76, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s87, v75, 31
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s86, v75, 30
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s85, v75, 29
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s84, v75, 28
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s83, v75, 27
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s82, v75, 26
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s81, v75, 25
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s80, v75, 24
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s71, v75, 23
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s70, v75, 22
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s69, v75, 21
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s68, v75, 20
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v75, 19
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v75, 18
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v75, 17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v75, 16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v75, 15
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v75, 14
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v75, 13
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v75, 12
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v75, 11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v75, 10
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v75, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v75, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v75, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v75, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v75, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v75, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v75, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v75, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v75, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v75, 0
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:88
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v32f32_to_v128i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:76
+; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:80
+; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:84
+; GFX11-NEXT:    scratch_store_b32 off, v78, s32 offset:88
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v75, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v76, s96, 0
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
+; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s41, v2
+; GFX11-NEXT:    v_writelane_b32 v75, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v76, s97, 1
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s15, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v5
+; GFX11-NEXT:    v_writelane_b32 v75, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v76, s98, 2
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v8
+; GFX11-NEXT:    v_writelane_b32 v75, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v76, s99, 3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v11
+; GFX11-NEXT:    v_writelane_b32 v75, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v76, s100, 4
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v13
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v14
+; GFX11-NEXT:    v_writelane_b32 v75, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v76, s101, 5
+; GFX11-NEXT:    s_mov_b32 vcc_hi, 0
+; GFX11-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
+; GFX11-NEXT:    s_clause 0x12
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:72
+; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:68
+; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:64
+; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:60
+; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:56
+; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:52
+; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:48
+; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:44
+; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:40
+; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:36
+; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:32
+; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:28
+; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:24
+; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:20
+; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:16
+; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:12
+; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v73, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v74, s32
+; GFX11-NEXT:    v_writelane_b32 v75, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v76, s102, 6
+; GFX11-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
+; GFX11-NEXT:    ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
+; GFX11-NEXT:    v_writelane_b32 v75, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v76, s103, 7
+; GFX11-NEXT:    v_writelane_b32 v75, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v76, s104, 8
+; GFX11-NEXT:    v_writelane_b32 v75, s49, 9
+; GFX11-NEXT:    v_writelane_b32 v75, s50, 10
+; GFX11-NEXT:    v_writelane_b32 v75, s51, 11
+; GFX11-NEXT:    v_writelane_b32 v75, s52, 12
+; GFX11-NEXT:    v_writelane_b32 v75, s53, 13
+; GFX11-NEXT:    v_writelane_b32 v75, s54, 14
+; GFX11-NEXT:    v_writelane_b32 v75, s55, 15
+; GFX11-NEXT:    v_writelane_b32 v75, s64, 16
+; GFX11-NEXT:    v_writelane_b32 v75, s65, 17
+; GFX11-NEXT:    v_writelane_b32 v75, s66, 18
+; GFX11-NEXT:    v_writelane_b32 v75, s67, 19
+; GFX11-NEXT:    v_writelane_b32 v75, s68, 20
+; GFX11-NEXT:    v_writelane_b32 v75, s69, 21
+; GFX11-NEXT:    v_writelane_b32 v75, s70, 22
+; GFX11-NEXT:    v_writelane_b32 v75, s71, 23
+; GFX11-NEXT:    v_writelane_b32 v75, s80, 24
+; GFX11-NEXT:    v_writelane_b32 v75, s81, 25
+; GFX11-NEXT:    v_writelane_b32 v75, s82, 26
+; GFX11-NEXT:    v_writelane_b32 v75, s83, 27
+; GFX11-NEXT:    v_writelane_b32 v75, s84, 28
+; GFX11-NEXT:    v_writelane_b32 v75, s85, 29
+; GFX11-NEXT:    v_writelane_b32 v75, s86, 30
+; GFX11-NEXT:    v_writelane_b32 v75, s87, 31
+; GFX11-NEXT:    s_cbranch_scc0 .LBB37_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 24
+; GFX11-NEXT:    s_lshr_b32 s36, s27, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s38, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s37, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s39, s26, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 7
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 8
+; GFX11-NEXT:    s_lshr_b32 s48, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s49, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s51, s25, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 6
+; GFX11-NEXT:    s_lshr_b32 s42, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s50, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s52, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s53, s23, 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 5
+; GFX11-NEXT:    s_lshr_b32 s42, s12, 8
+; GFX11-NEXT:    s_lshr_b32 s54, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s64, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s55, s22, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 4
+; GFX11-NEXT:    s_lshr_b32 s42, s11, 24
+; GFX11-NEXT:    s_lshr_b32 s65, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s66, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s67, s21, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s69, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s68, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s70, s20, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 2
+; GFX11-NEXT:    s_lshr_b32 s42, s11, 8
+; GFX11-NEXT:    s_lshr_b32 s71, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s80, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s82, s19, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 1
+; GFX11-NEXT:    s_lshr_b32 s42, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s81, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s83, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s84, s17, 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 0
+; GFX11-NEXT:    s_lshr_b32 s42, s10, 8
+; GFX11-NEXT:    s_lshr_b32 s85, s17, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 31
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 24
+; GFX11-NEXT:    s_lshr_b32 s87, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s86, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s96, s16, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 30
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s97, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s98, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s100, s3, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 29
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 8
+; GFX11-NEXT:    s_lshr_b32 s99, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s101, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s102, s1, 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 28
+; GFX11-NEXT:    s_lshr_b32 s42, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s103, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s34, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s104, s0, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 27
+; GFX11-NEXT:    s_lshr_b32 s42, s8, 8
+; GFX11-NEXT:    s_lshr_b32 s35, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
+; GFX11-NEXT:    s_lshr_b64 s[72:73], s[10:11], 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 26
+; GFX11-NEXT:    s_lshr_b32 s42, s7, 24
+; GFX11-NEXT:    s_lshr_b64 s[74:75], s[8:9], 24
+; GFX11-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
+; GFX11-NEXT:    s_lshr_b64 s[78:79], s[4:5], 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 25
+; GFX11-NEXT:    s_lshr_b32 s42, s7, 16
+; GFX11-NEXT:    s_lshr_b64 s[88:89], s[14:15], 24
+; GFX11-NEXT:    s_lshr_b64 s[90:91], s[40:41], 24
+; GFX11-NEXT:    s_lshr_b64 s[92:93], s[28:29], 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s7, 8
+; GFX11-NEXT:    s_lshr_b64 s[94:95], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 23
+; GFX11-NEXT:    s_lshr_b32 s42, s6, 16
+; GFX11-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 22
+; GFX11-NEXT:    s_lshr_b32 s42, s6, 8
+; GFX11-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 21
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 20
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 19
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 18
+; GFX11-NEXT:    s_lshr_b32 s42, s4, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 17
+; GFX11-NEXT:    s_lshr_b32 s42, s4, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s15, 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 15
+; GFX11-NEXT:    s_lshr_b32 s42, s15, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 14
+; GFX11-NEXT:    s_lshr_b32 s42, s15, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 13
+; GFX11-NEXT:    s_lshr_b32 s42, s14, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 12
+; GFX11-NEXT:    s_lshr_b32 s42, s14, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 11
+; GFX11-NEXT:    s_lshr_b32 s42, s41, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 10
+; GFX11-NEXT:    s_lshr_b32 s42, s41, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 9
+; GFX11-NEXT:    s_lshr_b32 s42, s41, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s40, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 7
+; GFX11-NEXT:    s_lshr_b32 s42, s40, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 6
+; GFX11-NEXT:    s_lshr_b32 s42, s29, 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 5
+; GFX11-NEXT:    s_lshr_b32 s42, s29, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 4
+; GFX11-NEXT:    s_lshr_b32 s42, s29, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 2
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 1
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 0
+; GFX11-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
+; GFX11-NEXT:    s_cbranch_vccnz .LBB37_4
+; GFX11-NEXT:  .LBB37_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f32_e64 v22, s27, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v21, s26, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v24, s25, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v23, s24, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v29, s23, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v28, s22, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v31, s21, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v30, s20, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v35, s19, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v34, s18, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v37, s17, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v36, s16, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v6, s9, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v5, s8, 1.0
+; GFX11-NEXT:    v_lshrrev_b64 v[64:65], 24, v[21:22]
+; GFX11-NEXT:    v_add_f32_e64 v53, s1, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v52, s0, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v49, s3, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v48, s2, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v18, s29, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v17, s28, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v14, s41, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v13, s40, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v12, s15, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v11, s14, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v10, s5, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v8, s7, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v4, s11, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v2, s13, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v1, s12, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v3, s10, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v7, s6, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v9, s4, 1.0
+; GFX11-NEXT:    v_lshrrev_b64 v[65:66], 24, v[23:24]
+; GFX11-NEXT:    v_lshrrev_b64 v[66:67], 24, v[28:29]
+; GFX11-NEXT:    v_lshrrev_b64 v[67:68], 24, v[30:31]
+; GFX11-NEXT:    v_lshrrev_b64 v[68:69], 24, v[34:35]
+; GFX11-NEXT:    v_lshrrev_b64 v[25:26], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b64 v[69:70], 24, v[36:37]
+; GFX11-NEXT:    v_lshrrev_b64 v[15:16], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[19:20], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[26:27], 24, v[7:8]
+; GFX11-NEXT:    v_lshrrev_b64 v[32:33], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[38:39], 24, v[11:12]
+; GFX11-NEXT:    v_lshrrev_b64 v[50:51], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]
+; GFX11-NEXT:    v_lshrrev_b64 v[70:71], 24, v[48:49]
+; GFX11-NEXT:    v_lshrrev_b64 v[80:81], 24, v[52:53]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 24, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 24, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 8, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v97, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v99, 24, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v98, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v100, 8, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v101, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v102, 8, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v112, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v103, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v113, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v114, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v115, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v117, 24, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v116, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v118, 8, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v119, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v128, 8, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v130, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v129, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v131, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v132, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v133, 8, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v135, 24, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v134, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v144, 8, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v145, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v146, 8, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v148, 24, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v147, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v149, 8, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v150, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v151, 8, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v160, 24, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v161, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v162, 8, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v163, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v164, 8, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v165, 24, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v166, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v167, 8, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v176, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v177, 8, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v178, 24, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v179, 16, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v180, 8, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v181, 16, v30
+; GFX11-NEXT:    v_lshrrev_b32_e32 v182, 8, v30
+; GFX11-NEXT:    v_lshrrev_b32_e32 v183, 24, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v40, 16, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v41, 8, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v42, 16, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v43, 8, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v44, 24, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v45, 16, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v46, 8, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v47, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v56, 8, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v57, 24, v49
+; GFX11-NEXT:    v_lshrrev_b32_e32 v58, 16, v49
+; GFX11-NEXT:    v_lshrrev_b32_e32 v59, 8, v49
+; GFX11-NEXT:    v_lshrrev_b32_e32 v60, 16, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v61, 8, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v62, 24, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v63, 16, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v72, 8, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v73, 16, v52
+; GFX11-NEXT:    v_lshrrev_b32_e32 v74, 8, v52
+; GFX11-NEXT:    s_branch .LBB37_5
+; GFX11-NEXT:  .LBB37_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr104
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr103
+; GFX11-NEXT:    ; implicit-def: $sgpr102
+; GFX11-NEXT:    ; implicit-def: $sgpr101
+; GFX11-NEXT:    ; implicit-def: $sgpr99
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr100
+; GFX11-NEXT:    ; implicit-def: $sgpr98
+; GFX11-NEXT:    ; implicit-def: $sgpr97
+; GFX11-NEXT:    ; implicit-def: $sgpr96
+; GFX11-NEXT:    ; implicit-def: $sgpr86
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr87
+; GFX11-NEXT:    ; implicit-def: $sgpr85
+; GFX11-NEXT:    ; implicit-def: $sgpr84
+; GFX11-NEXT:    ; implicit-def: $sgpr83
+; GFX11-NEXT:    ; implicit-def: $sgpr81
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr82
+; GFX11-NEXT:    ; implicit-def: $sgpr80
+; GFX11-NEXT:    ; implicit-def: $sgpr71
+; GFX11-NEXT:    ; implicit-def: $sgpr70
+; GFX11-NEXT:    ; implicit-def: $sgpr68
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr69
+; GFX11-NEXT:    ; implicit-def: $sgpr67
+; GFX11-NEXT:    ; implicit-def: $sgpr66
+; GFX11-NEXT:    ; implicit-def: $sgpr65
+; GFX11-NEXT:    ; implicit-def: $sgpr55
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr64
+; GFX11-NEXT:    ; implicit-def: $sgpr54
+; GFX11-NEXT:    ; implicit-def: $sgpr53
+; GFX11-NEXT:    ; implicit-def: $sgpr52
+; GFX11-NEXT:    ; implicit-def: $sgpr50
+; GFX11-NEXT:    ; implicit-def: $sgpr51
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    s_branch .LBB37_2
+; GFX11-NEXT:  .LBB37_4:
+; GFX11-NEXT:    v_dual_mov_b32 v52, s0 :: v_dual_mov_b32 v53, s1
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 0
+; GFX11-NEXT:    v_dual_mov_b32 v147, s36 :: v_dual_mov_b32 v48, s2
+; GFX11-NEXT:    v_dual_mov_b32 v49, s3 :: v_dual_mov_b32 v36, s16
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v37, s17 :: v_dual_mov_b32 v148, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 1
+; GFX11-NEXT:    v_dual_mov_b32 v34, s18 :: v_dual_mov_b32 v35, s19
+; GFX11-NEXT:    v_dual_mov_b32 v30, s20 :: v_dual_mov_b32 v31, s21
+; GFX11-NEXT:    v_mov_b32_e32 v146, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 2
+; GFX11-NEXT:    v_dual_mov_b32 v28, s22 :: v_dual_mov_b32 v29, s23
+; GFX11-NEXT:    v_dual_mov_b32 v23, s24 :: v_dual_mov_b32 v24, s25
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v145, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 3
+; GFX11-NEXT:    v_dual_mov_b32 v21, s26 :: v_dual_mov_b32 v22, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v18, s29
+; GFX11-NEXT:    v_mov_b32_e32 v144, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 4
+; GFX11-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
+; GFX11-NEXT:    v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v134, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 5
+; GFX11-NEXT:    v_dual_mov_b32 v9, s4 :: v_dual_mov_b32 v10, s5
+; GFX11-NEXT:    v_dual_mov_b32 v7, s6 :: v_dual_mov_b32 v8, s7
+; GFX11-NEXT:    v_mov_b32_e32 v135, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 6
+; GFX11-NEXT:    v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
+; GFX11-NEXT:    v_dual_mov_b32 v3, s10 :: v_dual_mov_b32 v4, s11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v133, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 7
+; GFX11-NEXT:    v_dual_mov_b32 v1, s12 :: v_dual_mov_b32 v2, s13
+; GFX11-NEXT:    v_dual_mov_b32 v74, s35 :: v_dual_mov_b32 v73, s104
+; GFX11-NEXT:    v_mov_b32_e32 v132, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 8
+; GFX11-NEXT:    v_dual_mov_b32 v72, s34 :: v_dual_mov_b32 v63, s103
+; GFX11-NEXT:    v_dual_mov_b32 v62, s102 :: v_dual_mov_b32 v61, s101
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v131, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 9
+; GFX11-NEXT:    v_dual_mov_b32 v60, s99 :: v_dual_mov_b32 v59, s100
+; GFX11-NEXT:    v_dual_mov_b32 v58, s98 :: v_dual_mov_b32 v57, s97
+; GFX11-NEXT:    v_mov_b32_e32 v129, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 10
+; GFX11-NEXT:    v_dual_mov_b32 v56, s96 :: v_dual_mov_b32 v47, s86
+; GFX11-NEXT:    v_dual_mov_b32 v46, s87 :: v_dual_mov_b32 v45, s85
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v130, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 11
+; GFX11-NEXT:    v_dual_mov_b32 v44, s84 :: v_dual_mov_b32 v43, s83
+; GFX11-NEXT:    v_dual_mov_b32 v42, s81 :: v_dual_mov_b32 v41, s82
+; GFX11-NEXT:    v_mov_b32_e32 v128, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 12
+; GFX11-NEXT:    v_dual_mov_b32 v40, s80 :: v_dual_mov_b32 v183, s71
+; GFX11-NEXT:    v_dual_mov_b32 v182, s70 :: v_dual_mov_b32 v181, s68
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v119, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 13
+; GFX11-NEXT:    v_dual_mov_b32 v180, s69 :: v_dual_mov_b32 v179, s67
+; GFX11-NEXT:    v_dual_mov_b32 v178, s66 :: v_dual_mov_b32 v177, s65
+; GFX11-NEXT:    v_mov_b32_e32 v118, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 14
+; GFX11-NEXT:    v_dual_mov_b32 v176, s55 :: v_dual_mov_b32 v167, s64
+; GFX11-NEXT:    v_dual_mov_b32 v166, s54 :: v_dual_mov_b32 v165, s53
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v116, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 15
+; GFX11-NEXT:    v_dual_mov_b32 v164, s52 :: v_dual_mov_b32 v163, s50
+; GFX11-NEXT:    v_dual_mov_b32 v162, s51 :: v_dual_mov_b32 v161, s49
+; GFX11-NEXT:    v_mov_b32_e32 v117, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 16
+; GFX11-NEXT:    v_dual_mov_b32 v160, s48 :: v_dual_mov_b32 v151, s39
+; GFX11-NEXT:    v_dual_mov_b32 v150, s37 :: v_dual_mov_b32 v149, s38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v115, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 17
+; GFX11-NEXT:    v_dual_mov_b32 v15, s62 :: v_dual_mov_b32 v38, s88
+; GFX11-NEXT:    v_dual_mov_b32 v19, s72 :: v_dual_mov_b32 v50, s90
+; GFX11-NEXT:    v_mov_b32_e32 v114, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 18
+; GFX11-NEXT:    v_dual_mov_b32 v25, s74 :: v_dual_mov_b32 v54, s92
+; GFX11-NEXT:    v_dual_mov_b32 v64, s94 :: v_dual_mov_b32 v65, s30
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v113, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 19
+; GFX11-NEXT:    v_dual_mov_b32 v66, s60 :: v_dual_mov_b32 v67, s58
+; GFX11-NEXT:    v_dual_mov_b32 v68, s56 :: v_dual_mov_b32 v69, s46
+; GFX11-NEXT:    v_mov_b32_e32 v103, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 20
+; GFX11-NEXT:    v_mov_b32_e32 v70, s44
+; GFX11-NEXT:    v_mov_b32_e32 v80, s42
+; GFX11-NEXT:    v_mov_b32_e32 v26, s76
+; GFX11-NEXT:    v_mov_b32_e32 v32, s78
+; GFX11-NEXT:    v_mov_b32_e32 v112, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 21
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v102, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 22
+; GFX11-NEXT:    v_mov_b32_e32 v101, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 23
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v100, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 24
+; GFX11-NEXT:    v_mov_b32_e32 v98, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 25
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v99, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 26
+; GFX11-NEXT:    v_mov_b32_e32 v97, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 27
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v96, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 28
+; GFX11-NEXT:    v_mov_b32_e32 v87, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 29
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v85, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 30
+; GFX11-NEXT:    v_mov_b32_e32 v86, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 31
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v84, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 0
+; GFX11-NEXT:    v_mov_b32_e32 v83, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v82, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 2
+; GFX11-NEXT:    v_mov_b32_e32 v51, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v55, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 4
+; GFX11-NEXT:    v_mov_b32_e32 v39, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v33, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 6
+; GFX11-NEXT:    v_mov_b32_e32 v27, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v20, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 8
+; GFX11-NEXT:    v_mov_b32_e32 v16, s0
+; GFX11-NEXT:  .LBB37_5: ; %end
+; GFX11-NEXT:    v_lshlrev_b32_e32 v71, 8, v80
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v74
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xff, v52
+; GFX11-NEXT:    v_and_b32_e32 v53, 0xff, v53
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v62
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xff, v48
+; GFX11-NEXT:    v_or_b32_e32 v52, v52, v80
+; GFX11-NEXT:    v_and_b32_e32 v80, 0xff, v73
+; GFX11-NEXT:    v_and_b32_e32 v60, 0xff, v60
+; GFX11-NEXT:    v_and_b32_e32 v49, 0xff, v49
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xff, v36
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v52
+; GFX11-NEXT:    v_or_b32_e32 v71, v80, v71
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v72
+; GFX11-NEXT:    v_or_b32_e32 v70, v60, v70
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v69
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xff, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v71, 16, v71
+; GFX11-NEXT:    v_or_b32_e32 v53, v53, v80
+; GFX11-NEXT:    v_and_b32_e32 v80, 0xff, v63
+; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xff, v30
+; GFX11-NEXT:    v_or_b32_e32 v60, v52, v71
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v53
+; GFX11-NEXT:    v_or_b32_e32 v80, v80, v81
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v61
+; GFX11-NEXT:    v_lshlrev_b32_e32 v71, 8, v59
+; GFX11-NEXT:    v_lshlrev_b32_e32 v67, 8, v67
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 16, v80
+; GFX11-NEXT:    v_or_b32_e32 v48, v48, v81
+; GFX11-NEXT:    v_and_b32_e32 v80, 0xff, v58
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v57
+; GFX11-NEXT:    v_or_b32_e32 v49, v49, v71
+; GFX11-NEXT:    v_or_b32_e32 v61, v52, v53
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v48
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 16, v70
+; GFX11-NEXT:    v_or_b32_e32 v53, v80, v81
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v56
+; GFX11-NEXT:    v_and_b32_e32 v71, 0xff, v47
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
+; GFX11-NEXT:    v_or_b32_e32 v62, v48, v52
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v49
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 16, v53
+; GFX11-NEXT:    v_or_b32_e32 v36, v36, v70
+; GFX11-NEXT:    v_or_b32_e32 v52, v71, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 8, v46
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xff, v45
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v44
+; GFX11-NEXT:    v_or_b32_e32 v63, v48, v49
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v52
+; GFX11-NEXT:    v_or_b32_e32 v37, v37, v53
+; GFX11-NEXT:    v_or_b32_e32 v49, v69, v70
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xff, v34
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 8, v43
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xff, v42
+; GFX11-NEXT:    v_or_b32_e32 v34, v36, v48
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 16, v49
+; GFX11-NEXT:    v_or_b32_e32 v48, v52, v53
+; GFX11-NEXT:    v_or_b32_e32 v49, v69, v68
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xff, v35
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 8, v41
+; GFX11-NEXT:    v_and_b32_e32 v68, 0xff, v40
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v183
+; GFX11-NEXT:    v_or_b32_e32 v35, v36, v37
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v48
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 16, v49
+; GFX11-NEXT:    v_or_b32_e32 v48, v52, v53
+; GFX11-NEXT:    v_or_b32_e32 v49, v68, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 8, v182
+; GFX11-NEXT:    v_and_b32_e32 v53, 0xff, v181
+; GFX11-NEXT:    v_or_b32_e32 v36, v36, v37
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v48
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
+; GFX11-NEXT:    v_or_b32_e32 v30, v30, v52
+; GFX11-NEXT:    v_or_b32_e32 v52, v53, v67
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 8, v177
+; GFX11-NEXT:    v_and_b32_e32 v67, 0xff, v176
+; GFX11-NEXT:    v_or_b32_e32 v37, v48, v49
+; GFX11-NEXT:    v_and_b32_e32 v49, 0xff, v179
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v52
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 8, v178
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xff, v31
+; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 8, v180
+; GFX11-NEXT:    v_or_b32_e32 v28, v28, v53
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xffff, v30
+; GFX11-NEXT:    v_or_b32_e32 v49, v49, v52
+; GFX11-NEXT:    v_or_b32_e32 v52, v67, v66
+; GFX11-NEXT:    v_or_b32_e32 v31, v31, v68
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 8, v167
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xff, v166
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v165
+; GFX11-NEXT:    v_and_b32_e32 v66, 0xffff, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_or_b32_e32 v29, v29, v68
+; GFX11-NEXT:    v_or_b32_e32 v53, v69, v70
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
+; GFX11-NEXT:    v_or_b32_e32 v28, v30, v48
+; GFX11-NEXT:    v_or_b32_e32 v30, v66, v52
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 8, v151
+; GFX11-NEXT:    v_and_b32_e32 v67, 0xffff, v29
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    scratch_store_b128 v0, v[60:63], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[34:37], off offset:16
+; GFX11-NEXT:    v_or_b32_e32 v29, v31, v49
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xff, v23
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 8, v164
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xff, v163
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 8, v65
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 8, v162
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xff, v161
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 8, v160
+; GFX11-NEXT:    v_or_b32_e32 v21, v21, v52
+; GFX11-NEXT:    v_or_b32_e32 v31, v67, v53
+; GFX11-NEXT:    v_or_b32_e32 v23, v23, v34
+; GFX11-NEXT:    v_or_b32_e32 v34, v35, v36
+; GFX11-NEXT:    v_or_b32_e32 v24, v24, v37
+; GFX11-NEXT:    v_or_b32_e32 v35, v48, v49
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v150
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 8, v64
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v149
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 8, v146
+; GFX11-NEXT:    v_and_b32_e32 v64, 0xff, v145
+; GFX11-NEXT:    v_lshlrev_b32_e32 v54, 8, v54
+; GFX11-NEXT:    v_and_b32_e32 v49, 0xff, v147
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 8, v148
+; GFX11-NEXT:    v_or_b32_e32 v21, v21, v37
+; GFX11-NEXT:    v_or_b32_e32 v22, v22, v48
+; GFX11-NEXT:    v_or_b32_e32 v17, v17, v53
+; GFX11-NEXT:    v_or_b32_e32 v48, v64, v54
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
+; GFX11-NEXT:    v_or_b32_e32 v37, v49, v52
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v22
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
+; GFX11-NEXT:    v_or_b32_e32 v21, v23, v34
+; GFX11-NEXT:    v_or_b32_e32 v22, v24, v35
+; GFX11-NEXT:    v_or_b32_e32 v23, v36, v49
+; GFX11-NEXT:    v_or_b32_e32 v34, v17, v48
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v18
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v144
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xff, v134
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 8, v135
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xff, v132
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 8, v50
+; GFX11-NEXT:    v_or_b32_e32 v24, v52, v37
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 8, v133
+; GFX11-NEXT:    v_or_b32_e32 v17, v17, v18
+; GFX11-NEXT:    v_or_b32_e32 v18, v35, v36
+; GFX11-NEXT:    v_or_b32_e32 v35, v48, v49
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v50, 8, v131
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 8, v130
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xff, v129
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v50
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v128
+; GFX11-NEXT:    v_and_b32_e32 v49, 0xff, v119
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v50, 8, v118
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xff, v116
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 8, v117
+; GFX11-NEXT:    v_or_b32_e32 v35, v35, v37
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v48
+; GFX11-NEXT:    v_or_b32_e32 v37, v49, v38
+; GFX11-NEXT:    v_or_b32_e32 v12, v12, v50
+; GFX11-NEXT:    v_or_b32_e32 v38, v52, v53
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v35
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 16, v37
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
+; GFX11-NEXT:    v_or_b32_e32 v35, v17, v18
+; GFX11-NEXT:    v_or_b32_e32 v36, v13, v36
+; GFX11-NEXT:    v_or_b32_e32 v37, v14, v48
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v115
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v114
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v32
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v113
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v102
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v49
+; GFX11-NEXT:    v_or_b32_e32 v12, v12, v38
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v103
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 8, v112
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v13
+; GFX11-NEXT:    v_or_b32_e32 v13, v14, v17
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v18
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v48
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v101
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v26
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v97
+; GFX11-NEXT:    v_and_b32_e32 v49, 0xff, v96
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
+; GFX11-NEXT:    v_or_b32_e32 v14, v32, v38
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v100
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v98
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 8, v99
+; GFX11-NEXT:    v_or_b32_e32 v17, v17, v18
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v48
+; GFX11-NEXT:    v_or_b32_e32 v25, v49, v25
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v26
+; GFX11-NEXT:    v_or_b32_e32 v18, v32, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
+; GFX11-NEXT:    v_or_b32_e32 v13, v9, v13
+; GFX11-NEXT:    v_or_b32_e32 v14, v10, v14
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v17
+; GFX11-NEXT:    v_or_b32_e32 v9, v5, v25
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 8, v87
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v85
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v86
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v83
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v18
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v84
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v82
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v10, v17
+; GFX11-NEXT:    v_or_b32_e32 v10, v25, v19
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v18
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v26
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v55
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v51
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v39
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v33
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v27
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v18
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v19
+; GFX11-NEXT:    v_or_b32_e32 v15, v25, v15
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v26
+; GFX11-NEXT:    v_or_b32_e32 v16, v20, v16
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_or_b32_e32 v10, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v1, v3, v17
+; GFX11-NEXT:    v_or_b32_e32 v2, v4, v18
+; GFX11-NEXT:    v_or_b32_e32 v3, v19, v15
+; GFX11-NEXT:    v_or_b32_e32 v4, v20, v16
+; GFX11-NEXT:    s_clause 0x5
+; GFX11-NEXT:    scratch_store_b128 v0, v[28:31], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[21:24], off offset:48
+; GFX11-NEXT:    scratch_store_b128 v0, v[34:37], off offset:64
+; GFX11-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
+; GFX11-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
+; GFX11-NEXT:    s_clause 0x12
+; GFX11-NEXT:    scratch_load_b32 v74, off, s32
+; GFX11-NEXT:    scratch_load_b32 v73, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:12
+; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:16
+; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:20
+; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:24
+; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:28
+; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:32
+; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:36
+; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:40
+; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:44
+; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:48
+; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:52
+; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:56
+; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:60
+; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:64
+; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:68
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:72
+; GFX11-NEXT:    v_readlane_b32 s104, v76, 8
+; GFX11-NEXT:    v_readlane_b32 s103, v76, 7
+; GFX11-NEXT:    v_readlane_b32 s102, v76, 6
+; GFX11-NEXT:    v_readlane_b32 s101, v76, 5
+; GFX11-NEXT:    v_readlane_b32 s100, v76, 4
+; GFX11-NEXT:    v_readlane_b32 s99, v76, 3
+; GFX11-NEXT:    v_readlane_b32 s98, v76, 2
+; GFX11-NEXT:    v_readlane_b32 s97, v76, 1
+; GFX11-NEXT:    v_readlane_b32 s96, v76, 0
+; GFX11-NEXT:    v_readlane_b32 s87, v75, 31
+; GFX11-NEXT:    v_readlane_b32 s86, v75, 30
+; GFX11-NEXT:    v_readlane_b32 s85, v75, 29
+; GFX11-NEXT:    v_readlane_b32 s84, v75, 28
+; GFX11-NEXT:    v_readlane_b32 s83, v75, 27
+; GFX11-NEXT:    v_readlane_b32 s82, v75, 26
+; GFX11-NEXT:    v_readlane_b32 s81, v75, 25
+; GFX11-NEXT:    v_readlane_b32 s80, v75, 24
+; GFX11-NEXT:    v_readlane_b32 s71, v75, 23
+; GFX11-NEXT:    v_readlane_b32 s70, v75, 22
+; GFX11-NEXT:    v_readlane_b32 s69, v75, 21
+; GFX11-NEXT:    v_readlane_b32 s68, v75, 20
+; GFX11-NEXT:    v_readlane_b32 s67, v75, 19
+; GFX11-NEXT:    v_readlane_b32 s66, v75, 18
+; GFX11-NEXT:    v_readlane_b32 s65, v75, 17
+; GFX11-NEXT:    v_readlane_b32 s64, v75, 16
+; GFX11-NEXT:    v_readlane_b32 s55, v75, 15
+; GFX11-NEXT:    v_readlane_b32 s54, v75, 14
+; GFX11-NEXT:    v_readlane_b32 s53, v75, 13
+; GFX11-NEXT:    v_readlane_b32 s52, v75, 12
+; GFX11-NEXT:    v_readlane_b32 s51, v75, 11
+; GFX11-NEXT:    v_readlane_b32 s50, v75, 10
+; GFX11-NEXT:    v_readlane_b32 s49, v75, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v75, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v75, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v75, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v75, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v75, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v75, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v75, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v75, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v75, 0
+; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:76
+; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:80
+; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:84
+; GFX11-NEXT:    scratch_load_b32 v78, off, s32 offset:88
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -60364,462 +57402,482 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v32f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:436
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:320
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v86, v0
+; GFX11-TRUE16-NEXT:    s_clause 0x1f
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:476
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:472
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:468
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:464
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:460
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:456
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:452
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:448
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:444
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:352
+; GFX11-TRUE16-NEXT:    s_clause 0x7
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:320
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v26 :: v_dual_mov_b32 v35, v24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v22 :: v_dual_mov_b32 v37, v20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v18 :: v_dual_mov_b32 v39, v16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v14 :: v_dual_mov_b32 v49, v12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v10 :: v_dual_mov_b32 v51, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v6 :: v_dual_mov_b32 v53, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, v2 :: v_dual_mov_b32 v55, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
 ; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:316
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:312
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:308
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:304
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:300
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:296
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:292
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v2, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v16, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v18, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v20, off, s32 offset:72
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v22, off, s32 offset:80
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v24, off, s32 offset:88
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v26, off, s32 offset:96
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v28, off, s32 offset:104
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v30, off, s32 offset:112
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:120
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:128
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:136
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:144
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:152
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:160
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:168
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:176
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:184
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:200
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:208
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:216
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:224
 ; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v94, off, s32 offset:240
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v95, off, s32 offset:248
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v104, off, s32 offset:256
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v105, off, s32 offset:264
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v106, off, s32 offset:272
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v107, off, s32 offset:280
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v108, off, s32 offset:288
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v109, off, s32 offset:296
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v110, off, s32 offset:304
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v111, off, s32 offset:312
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:308
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:300
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:292
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:284
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:276
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:268
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:260
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:252
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:244
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:228
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:220
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:212
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:204
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:196
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:188
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:180
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:172
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:164
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:156
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:148
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:140
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:132
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v89, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v88, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v90, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v91, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v93, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v92, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v28 :: v_dual_mov_b32 v34, v27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v26 :: v_dual_mov_b32 v37, v25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v24 :: v_dual_mov_b32 v38, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v22 :: v_dual_mov_b32 v49, v21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v20 :: v_dual_mov_b32 v51, v19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v18 :: v_dual_mov_b32 v53, v17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v55, v16 :: v_dual_mov_b32 v54, v15
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, v14 :: v_dual_mov_b32 v65, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v67, v12 :: v_dual_mov_b32 v66, v11
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v68, v10 :: v_dual_mov_b32 v69, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v71, v8 :: v_dual_mov_b32 v70, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v80, v6 :: v_dual_mov_b32 v81, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v83, v4 :: v_dual_mov_b32 v82, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v84, v2 :: v_dual_mov_b32 v85, v1
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:124
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:116
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:108
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:100
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:92
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:84
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:76
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:68
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v89, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v90, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v91, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v92, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v93, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v76, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v77, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v78, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v79, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v88, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v63, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v72, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v73, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v74, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v75, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v57, 8, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v58, 8, v4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v59, 8, v6
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v60, 8, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v61, 8, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v40, 8, v12
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 8, v14
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v42, 8, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v43, 8, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v44, 8, v20
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v167, 8, v22
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v176, 8, v24
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v177, 8, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v178, 8, v28
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v179, 8, v30
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v149, 8, v31
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v150, 8, v87
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(61)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v151, 8, v96
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(60)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v160, 8, v97
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(59)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v161, 8, v98
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(58)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v132, 8, v99
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(57)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v133, 8, v113
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(56)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v134, 8, v114
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(55)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v135, 8, v115
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(54)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v144, 8, v116
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(53)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v119, 8, v117
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(52)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v128, 8, v128
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(51)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v129, 8, v129
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(50)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v130, 8, v130
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(49)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v131, 8, v131
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(48)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v113, 8, v94
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(47)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v114, 8, v95
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(46)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v115, 8, v104
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(45)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v116, 8, v105
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(44)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v117, 8, v106
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(43)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v107
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(42)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v108
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(41)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v109
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(40)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v98, 8, v110
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(39)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v99, 8, v111
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB39_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v71
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v53
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v90
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v91
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v68
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v65
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v67
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v53
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s26, 0xff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v49
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s27, 8
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v76
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v77
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v64
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v78
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v79
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v49
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v51
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v37
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v88
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v63
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v37
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v72
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v73
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v33
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v33
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v74
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v75
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v91
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v0, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v32
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v92
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v93
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v62
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v57
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v58
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v79
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v90
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v88
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v89
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v56
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v47
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v59
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v60
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v75
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v78
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v77
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v76
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v45
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v61
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v40
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v63
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v74
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v72
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v73
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v183
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v182
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v41
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v42
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v59
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v62
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v60
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v61
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v43
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v44
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v47
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v56
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v57
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v165
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v167
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v176
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v43
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v44
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v45
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v164
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v163
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v177
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v178
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v183
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v42
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v41
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v40
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v148
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v179
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v149
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v179
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v182
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v147
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v146
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v150
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v151
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v167
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v178
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v176
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v177
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v118
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v160
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v161
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v163
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v164
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v165
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v103
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v132
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v133
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v151
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v160
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v161
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v102
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v101
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v134
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v135
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v147
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v150
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v149
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v148
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v100
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v86
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v144
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v119
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v135
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v146
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v144
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v85
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v84
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v128
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v129
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v131
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v134
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v132
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v133
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v83
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v130
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v131
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v119
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v130
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v128
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v129
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v81
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v80
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v113
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v114
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v114
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v118
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v116
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v117
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v115
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v116
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v102
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v115
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v113
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v68
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v117
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v87
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v98
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v103
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v100
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v67
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v96
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v97
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v87
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v99
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v96
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v97
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v64
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v98
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v99
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v86
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v85
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v89
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
@@ -60852,8 +57910,18 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s10, 16
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v51
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v93
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v92
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB39_3
@@ -60871,445 +57939,401 @@ define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a,
 ; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s4
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
 ; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v86
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v55
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v54
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v52
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v85
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v51
 ; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v84
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v81
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v82
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v38
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v89, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v90, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v91, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v92, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v93, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v69
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v88, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v74, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v75, v12
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v68
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v65
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v64
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v8, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v54
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v49
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v48
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v39
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v37
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v8, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v46
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v181
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v76, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v77, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v78, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v79, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v63, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v72, v8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v73, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v49
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v35
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v9, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v33
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v3, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v93
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v61, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v43, v16
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v44, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v14, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v32
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v92
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v62
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v56
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v47
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v45
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v183
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v182
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v91
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v90
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v89
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v12, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v13, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v88
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v79
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v78
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v162
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v145
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v118
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v57, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v58, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v59, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v60, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v40, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v41, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v42, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v12, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v13, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v77
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v76
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v75
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v74
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v73
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v14, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v72
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v63
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v62
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v61
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v14, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v15, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v59
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v17, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v15
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v58
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v3, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v57
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v56
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v179, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v160, v21
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v161, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v18
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v19, v15
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v166
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v165
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v164
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v163
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v148
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v147
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v146
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v47
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v46
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v45
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v17, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v18, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v44
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v43
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v42
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v100
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v83
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v167, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v176, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v177, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v178, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v149, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v150, v18
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v151, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v17, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v18, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v41
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v40
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v183
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v182
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v181
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v19, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v180
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v178
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v177
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v19, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v20, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v176
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v167
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v22, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v166
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v3, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v165
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v164
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v144, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v130, v26
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v131, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v19, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v24, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v112
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v103
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v102
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v86
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v85
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v163
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v162
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v161
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v22, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v23, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v160
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v151
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v150
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v26, v27
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v69
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v132, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v133, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v134, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v135, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v119, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v128, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v129, v25
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v22, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v23, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v149
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v148
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v147
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v146
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v145
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v24, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v144
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v135
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v134
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v133
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v24, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v25, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v132
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v131
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v27, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v130
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v3, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v128
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v129
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v119
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v113
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v27, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v115
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v117, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v24, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v29, v25
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v81
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v80
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v71
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v70
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v27
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v68
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v67
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v66
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v65
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 3, v64
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v118
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v117
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v114
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v28, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v116
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v30, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v101
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 3, v99
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v28, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v29, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v112
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v103
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v97
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v102
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xff, v30
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v98
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v87
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v28, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v33, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v35, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v36, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v37, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v113, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v114, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v115, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v116, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v87, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v96, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v97, v30
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v98, v31
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v99, v32
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v27
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v2, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v1, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v30, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v34, v35
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v29, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v34, v30
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v31, v32
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-TRUE16-NEXT:  .LBB39_3: ; %end
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:320
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    s_clause 0x1f
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:320
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:352
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:444
+; GFX11-TRUE16-NEXT:    s_clause 0x7
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:448
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:452
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:456
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:460
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:464
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:468
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:472
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:476
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ; GFX11-TRUE16-NEXT:  .LBB39_4:
@@ -68870,1735 +65894,870 @@ define inreg <32 x float> @bitcast_v64bf16_to_v32f32_scalar(<64 x bfloat> inreg
 ; GFX9-NEXT:  .LBB43_4:
 ; GFX9-NEXT:    s_branch .LBB43_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v32f32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:192
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:164
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v120, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v121, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v122, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v123, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v124, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v125, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v126, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v127, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v136, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v137, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v138, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v139, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v140, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v141, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v142, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v143, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v152, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v153, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v154, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v155, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v156, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v157, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v158, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v159, s32 offset:36
-; GFX11-TRUE16-NEXT:    s_clause 0x8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v168, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v169, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v170, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v171, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v172, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v173, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v174, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v175, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v184, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v183, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v182, v9 :: v_dual_mov_b32 v169, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v171, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v176, v3 :: v_dual_mov_b32 v175, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v173, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v181, s28 :: v_dual_mov_b32 v172, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB43_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB43_3
-; GFX11-TRUE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s27, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s27, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s26, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s25, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s25, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff0000, v181
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v8, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v8
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v10, v6
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s24, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s23, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v7, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s22, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v11, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s22, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v9, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s21, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v14, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v11
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v16, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s20, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v12, v16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s20, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s19, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v18, v12
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v16, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s18, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v21, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v18
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s17, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v17, v20
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v19, v22
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s17, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v19, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s16, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, v24, v19
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v25, 16, 1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s16, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, v22, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v19
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, 0x400000, v25
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s3, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, v27, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v23
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v26, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v28, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s2, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, v25, v28
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v30, v25, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s1, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, v30, v25
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v31, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, v28, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v31
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v30, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v33, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, 0x400000, v29
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v30
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v29, v32, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v31, v34, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, v29, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, v31, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v31, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v37, v33
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v35, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v33, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v36, v37
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v183
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v35, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v183
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v37, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v182
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v182
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v35, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v36, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v183, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v38, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v39, v36
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v182, v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v177
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v35, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v177
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v32, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v49, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v49, v35
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v37, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v31, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v176
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v31, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v177, v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v50, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v176
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v49, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v50, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v37, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v48, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, v39, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v176, v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v34, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v37, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v39, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, v38, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v181
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v36, v49, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v36, v38
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, 0x400000, v48
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v50, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v39, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, 0x400000, v39
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v51, v48, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, v38, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v37
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v52, v50, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, v51, v48
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v52, v52, v50
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v52, 0x400000, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v181, v39, 16, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
-; GFX11-TRUE16-NEXT:  .LBB43_3: ; %end
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v181 :: v_dual_mov_b32 v19, v175
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v21, v176
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v184, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v175, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v174, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v173, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v172, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v171, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v170, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v169, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v168, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v159, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v158, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v157, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v156, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v155, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v154, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v153, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v152, off, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v143, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v142, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v141, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v140, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v139, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v138, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v137, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v136, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v127, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v126, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v125, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v124, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v123, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v122, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v121, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v120, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:192
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    s_clause 0x8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v177 :: v_dual_mov_b32 v27, v182
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v183 :: v_dual_mov_b32 v31, v178
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v30, v179
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB43_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
-; GFX11-TRUE16-NEXT:    s_branch .LBB43_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v32f32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:288
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:284
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:280
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:276
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:272
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:268
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:264
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:260
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:256
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:252
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:248
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:244
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:240
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:236
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:232
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:228
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:224
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:220
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:216
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:212
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:208
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:204
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:200
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:196
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:192
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:188
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:184
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:180
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:176
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:172
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:168
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:164
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:160
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:156
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:152
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:148
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:144
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:140
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:136
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:132
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:128
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:124
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:120
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:116
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:112
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:108
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:104
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:100
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v136, s32 offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v137, s32 offset:92
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v138, s32 offset:88
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v139, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v140, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v141, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v142, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v143, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v152, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v153, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v154, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v155, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v156, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v157, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v158, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v159, s32 offset:36
-; GFX11-FAKE16-NEXT:    s_clause 0x8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v168, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v169, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v170, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v171, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v172, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v173, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v174, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v175, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v184, s32
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v181, v9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v182, v10 :: v_dual_mov_b32 v169, v7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v176, v6 :: v_dual_mov_b32 v171, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v173, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v175, v1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v183, s28 :: v_dual_mov_b32 v172, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB43_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB43_3
-; GFX11-FAKE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s27, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s27, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s26, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s25, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s25, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s24, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff0000, v183
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v8, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v3, v8
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v10, v6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s24, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s23, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v7, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s23, 16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s22, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v11, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s22, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v9, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v10, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s21, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v14, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s21, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v11
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v16, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s20, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, v12, v16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s20, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s19, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v18, v12
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s19, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v16, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s18, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s18, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v18
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s17, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v17, v20
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v19, v22
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s17, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v19, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s16, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v24, v19
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v25, 16, 1
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v22, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v19
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v25
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s3, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, v27, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v23
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v26, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v25, v28, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s2, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, v25, v28
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, 0x400000, v28
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v30, v25, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s1, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, v30, v25
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v31, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, v28, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v31
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v30, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v33, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, 0x400000, v29
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, 0x400000, v30
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v29, v32, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v31, v34, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, v29, v32
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, v31, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v31, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v37, v33
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v35, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v33, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v36, v37
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v182
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v35, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v182
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v37, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v181
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v181
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v35, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v36, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v182, v31, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v38, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v39, v36
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v181, v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v176
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v35, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v176
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v32, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v49, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v37, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v31, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v177
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v31, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v176, v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v177
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v50, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v37, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v48, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, v39, v38
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v177, v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v34, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v37, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v39, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, v38, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v36, v49 :: v_dual_lshlrev_b32 v48, 16, v183
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v36, v38
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v48
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v39, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v39
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v48, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, v38, v39
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v37
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v50, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v51, v51, v48
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v52, v52, v50
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, 0x400000, v50
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v183, v39, 16, v48
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
-; GFX11-FAKE16-NEXT:  .LBB43_3: ; %end
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v19, v175
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v184, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v175, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v174, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v173, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v172, off, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v171, off, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v170, off, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v169, off, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v168, off, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v159, off, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v158, off, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v157, off, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v156, off, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v155, off, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v154, off, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v153, off, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v152, off, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v143, off, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v142, off, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v141, off, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v140, off, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v139, off, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v138, off, s32 offset:88
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v137, off, s32 offset:92
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v136, off, s32 offset:96
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:100
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:104
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:108
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:112
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:116
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:120
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:124
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:128
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:132
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:136
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:140
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:144
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:148
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:152
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:156
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:160
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:164
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:168
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:172
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:176
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:180
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:184
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:188
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:192
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:196
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:200
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:204
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:208
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:212
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:216
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:220
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:224
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:228
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:232
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:236
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:240
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:244
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:248
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:252
-; GFX11-FAKE16-NEXT:    s_clause 0x8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:256
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:260
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:264
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:268
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:272
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:276
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:280
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:284
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:288
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v183 :: v_dual_mov_b32 v21, v177
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, v176 :: v_dual_mov_b32 v27, v181
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v28, v182
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, v179 :: v_dual_mov_b32 v31, v178
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB43_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
-; GFX11-FAKE16-NEXT:    s_branch .LBB43_2
+; GFX11-LABEL: bitcast_v64bf16_to_v32f32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:288
+; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:284
+; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:280
+; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:276
+; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:272
+; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:268
+; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:264
+; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:260
+; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:256
+; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:252
+; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:248
+; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:244
+; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:240
+; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:236
+; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:232
+; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:228
+; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:224
+; GFX11-NEXT:    scratch_store_b32 off, v73, s32 offset:220
+; GFX11-NEXT:    scratch_store_b32 off, v74, s32 offset:216
+; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:212
+; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:208
+; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:204
+; GFX11-NEXT:    scratch_store_b32 off, v78, s32 offset:200
+; GFX11-NEXT:    scratch_store_b32 off, v79, s32 offset:196
+; GFX11-NEXT:    scratch_store_b32 off, v88, s32 offset:192
+; GFX11-NEXT:    scratch_store_b32 off, v89, s32 offset:188
+; GFX11-NEXT:    scratch_store_b32 off, v90, s32 offset:184
+; GFX11-NEXT:    scratch_store_b32 off, v91, s32 offset:180
+; GFX11-NEXT:    scratch_store_b32 off, v92, s32 offset:176
+; GFX11-NEXT:    scratch_store_b32 off, v93, s32 offset:172
+; GFX11-NEXT:    scratch_store_b32 off, v94, s32 offset:168
+; GFX11-NEXT:    scratch_store_b32 off, v95, s32 offset:164
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_store_b32 off, v104, s32 offset:160
+; GFX11-NEXT:    scratch_store_b32 off, v105, s32 offset:156
+; GFX11-NEXT:    scratch_store_b32 off, v106, s32 offset:152
+; GFX11-NEXT:    scratch_store_b32 off, v107, s32 offset:148
+; GFX11-NEXT:    scratch_store_b32 off, v108, s32 offset:144
+; GFX11-NEXT:    scratch_store_b32 off, v109, s32 offset:140
+; GFX11-NEXT:    scratch_store_b32 off, v110, s32 offset:136
+; GFX11-NEXT:    scratch_store_b32 off, v111, s32 offset:132
+; GFX11-NEXT:    scratch_store_b32 off, v120, s32 offset:128
+; GFX11-NEXT:    scratch_store_b32 off, v121, s32 offset:124
+; GFX11-NEXT:    scratch_store_b32 off, v122, s32 offset:120
+; GFX11-NEXT:    scratch_store_b32 off, v123, s32 offset:116
+; GFX11-NEXT:    scratch_store_b32 off, v124, s32 offset:112
+; GFX11-NEXT:    scratch_store_b32 off, v125, s32 offset:108
+; GFX11-NEXT:    scratch_store_b32 off, v126, s32 offset:104
+; GFX11-NEXT:    scratch_store_b32 off, v127, s32 offset:100
+; GFX11-NEXT:    scratch_store_b32 off, v136, s32 offset:96
+; GFX11-NEXT:    scratch_store_b32 off, v137, s32 offset:92
+; GFX11-NEXT:    scratch_store_b32 off, v138, s32 offset:88
+; GFX11-NEXT:    scratch_store_b32 off, v139, s32 offset:84
+; GFX11-NEXT:    scratch_store_b32 off, v140, s32 offset:80
+; GFX11-NEXT:    scratch_store_b32 off, v141, s32 offset:76
+; GFX11-NEXT:    scratch_store_b32 off, v142, s32 offset:72
+; GFX11-NEXT:    scratch_store_b32 off, v143, s32 offset:68
+; GFX11-NEXT:    scratch_store_b32 off, v152, s32 offset:64
+; GFX11-NEXT:    scratch_store_b32 off, v153, s32 offset:60
+; GFX11-NEXT:    scratch_store_b32 off, v154, s32 offset:56
+; GFX11-NEXT:    scratch_store_b32 off, v155, s32 offset:52
+; GFX11-NEXT:    scratch_store_b32 off, v156, s32 offset:48
+; GFX11-NEXT:    scratch_store_b32 off, v157, s32 offset:44
+; GFX11-NEXT:    scratch_store_b32 off, v158, s32 offset:40
+; GFX11-NEXT:    scratch_store_b32 off, v159, s32 offset:36
+; GFX11-NEXT:    s_clause 0x8
+; GFX11-NEXT:    scratch_store_b32 off, v168, s32 offset:32
+; GFX11-NEXT:    scratch_store_b32 off, v169, s32 offset:28
+; GFX11-NEXT:    scratch_store_b32 off, v170, s32 offset:24
+; GFX11-NEXT:    scratch_store_b32 off, v171, s32 offset:20
+; GFX11-NEXT:    scratch_store_b32 off, v172, s32 offset:16
+; GFX11-NEXT:    scratch_store_b32 off, v173, s32 offset:12
+; GFX11-NEXT:    scratch_store_b32 off, v174, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v175, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v184, s32
+; GFX11-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
+; GFX11-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v181, v9
+; GFX11-NEXT:    v_dual_mov_b32 v182, v10 :: v_dual_mov_b32 v169, v7
+; GFX11-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v3
+; GFX11-NEXT:    v_dual_mov_b32 v176, v6 :: v_dual_mov_b32 v171, v4
+; GFX11-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v173, v0
+; GFX11-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v175, v1
+; GFX11-NEXT:    v_dual_mov_b32 v183, s28 :: v_dual_mov_b32 v172, s29
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB43_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
+; GFX11-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
+; GFX11-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
+; GFX11-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
+; GFX11-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
+; GFX11-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
+; GFX11-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
+; GFX11-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB43_3
+; GFX11-NEXT:  .LBB43_2: ; %cmp.true
+; GFX11-NEXT:    s_and_b32 s5, s27, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s4, s27, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s6, s26, 16
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v1
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_bfe_u32 v10, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_bfe_u32 v9, v3, 16, 1
+; GFX11-NEXT:    s_lshl_b32 s7, s25, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    s_and_b32 s5, s25, 0xffff0000
+; GFX11-NEXT:    s_and_b32 s4, s24, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
+; GFX11-NEXT:    v_and_b32_e32 v51, 0xffff0000, v183
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v10, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v9, v3
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_bfe_u32 v10, v6, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_bfe_u32 v3, v8, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v3, v8
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v10, v6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    s_lshl_b32 s4, s24, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v6
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v8
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
+; GFX11-NEXT:    s_and_b32 s4, s23, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v4, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v5
+; GFX11-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
+; GFX11-NEXT:    v_bfe_u32 v7, v9, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, v7, v9
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s23, 16
+; GFX11-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_bfe_u32 v11, v7, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    s_and_b32 s4, s22, 0xffff0000
+; GFX11-NEXT:    v_bfe_u32 v9, v12, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v11, v7
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s22, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
+; GFX11-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v9, v12
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v7
+; GFX11-NEXT:    v_bfe_u32 v14, v10, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_or_b32_e32 v13, 0x400000, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
+; GFX11-NEXT:    s_and_b32 s4, s21, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v9, v11, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, v14, v10
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
+; GFX11-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v10
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    s_lshl_b32 s4, s21, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v11
+; GFX11-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v10, v13, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_bfe_u32 v12, v16, 16, 1
+; GFX11-NEXT:    s_and_b32 s4, s20, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, v12, v16
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    s_lshl_b32 s4, s20, 16
+; GFX11-NEXT:    v_or_b32_e32 v17, 0x400000, v16
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
+; GFX11-NEXT:    v_bfe_u32 v18, v12, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
+; GFX11-NEXT:    s_and_b32 s4, s19, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v18, v12
+; GFX11-NEXT:    v_bfe_u32 v16, v19, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v9
+; GFX11-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s19, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v16, v19
+; GFX11-NEXT:    v_or_b32_e32 v16, 0x400000, v12
+; GFX11-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v21, v17, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
+; GFX11-NEXT:    v_or_b32_e32 v20, 0x400000, v19
+; GFX11-NEXT:    s_and_b32 s4, s18, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v16, v18, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
+; GFX11-NEXT:    v_add_nc_u32_e32 v19, v21, v17
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
+; GFX11-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
+; GFX11-NEXT:    v_or_b32_e32 v19, 0x400000, v17
+; GFX11-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
+; GFX11-NEXT:    s_lshl_b32 s4, s18, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
+; GFX11-NEXT:    v_or_b32_e32 v21, 0x400000, v18
+; GFX11-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v17, v20, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_bfe_u32 v19, v22, 16, 1
+; GFX11-NEXT:    s_and_b32 s4, s17, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v17, v17, v20
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
+; GFX11-NEXT:    v_or_b32_e32 v21, 0x400000, v20
+; GFX11-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
+; GFX11-NEXT:    v_add_nc_u32_e32 v18, v19, v22
+; GFX11-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
+; GFX11-NEXT:    s_lshl_b32 s4, s17, 16
+; GFX11-NEXT:    v_or_b32_e32 v23, 0x400000, v22
+; GFX11-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
+; GFX11-NEXT:    v_bfe_u32 v24, v19, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
+; GFX11-NEXT:    s_and_b32 s4, s16, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, v24, v19
+; GFX11-NEXT:    v_bfe_u32 v22, v25, 16, 1
+; GFX11-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
+; GFX11-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s16, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
+; GFX11-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, v22, v25
+; GFX11-NEXT:    v_or_b32_e32 v22, 0x400000, v19
+; GFX11-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v27, v23, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
+; GFX11-NEXT:    v_or_b32_e32 v26, 0x400000, v25
+; GFX11-NEXT:    s_and_b32 s4, s3, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v22, v24, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
+; GFX11-NEXT:    v_add_nc_u32_e32 v25, v27, v23
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
+; GFX11-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
+; GFX11-NEXT:    v_or_b32_e32 v25, 0x400000, v23
+; GFX11-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
+; GFX11-NEXT:    v_or_b32_e32 v27, 0x400000, v24
+; GFX11-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
+; GFX11-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v23, v26, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
+; GFX11-NEXT:    v_bfe_u32 v25, v28, 16, 1
+; GFX11-NEXT:    s_and_b32 s3, s2, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
+; GFX11-NEXT:    v_or_b32_e32 v27, 0x400000, v26
+; GFX11-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v24, v25, v28
+; GFX11-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_or_b32_e32 v29, 0x400000, v28
+; GFX11-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
+; GFX11-NEXT:    v_bfe_u32 v30, v25, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
+; GFX11-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
+; GFX11-NEXT:    s_and_b32 s2, s1, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, v30, v25
+; GFX11-NEXT:    v_bfe_u32 v28, v31, 16, 1
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, v28, v31
+; GFX11-NEXT:    v_or_b32_e32 v28, 0x400000, v25
+; GFX11-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
+; GFX11-NEXT:    v_bfe_u32 v33, v29, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
+; GFX11-NEXT:    v_or_b32_e32 v32, 0x400000, v31
+; GFX11-NEXT:    s_and_b32 s1, s0, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v28, v30, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v33, v29
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
+; GFX11-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
+; GFX11-NEXT:    v_or_b32_e32 v31, 0x400000, v29
+; GFX11-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
+; GFX11-NEXT:    v_or_b32_e32 v33, 0x400000, v30
+; GFX11-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
+; GFX11-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v29, v32, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
+; GFX11-NEXT:    v_or_b32_e32 v35, 0x400000, v32
+; GFX11-NEXT:    v_bfe_u32 v31, v34, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
+; GFX11-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, v29, v32
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
+; GFX11-NEXT:    v_add_nc_u32_e32 v30, v31, v34
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
+; GFX11-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
+; GFX11-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
+; GFX11-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v37, v33, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v29
+; GFX11-NEXT:    v_bfe_u32 v35, v31, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v37, v33
+; GFX11-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v35, v31
+; GFX11-NEXT:    v_or_b32_e32 v37, 0x400000, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_or_b32_e32 v38, 0x400000, v31
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v34, v36, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v33, v35, 16, 1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v33, v35
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_bfe_u32 v36, v37, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_bfe_u32 v35, v38, 16, 1
+; GFX11-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v36, v37
+; GFX11-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v182
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v35, v38
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v182
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xffff, v30
+; GFX11-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_bfe_u32 v37, v36, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v37, v36
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v181
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v181
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-NEXT:    v_bfe_u32 v35, v37, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v36, v38, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v35, v37
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v36, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v37
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v39, v36, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_lshl_or_b32 v182, v31, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v38, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v39, v36
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v38, 0x400000, v36
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
+; GFX11-NEXT:    v_lshl_or_b32 v181, v32, 16, v33
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff0000, v176
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_bfe_u32 v35, v37, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
+; GFX11-NEXT:    v_bfe_u32 v32, v36, 16, 1
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v35, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 16, v176
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v32, v36
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_bfe_u32 v37, v38, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v49, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v49, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v37, v35, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
+; GFX11-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v37, v35
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v31, v36, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v177
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v31, v36
+; GFX11-NEXT:    v_lshl_or_b32 v176, v33, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
+; GFX11-NEXT:    v_or_b32_e32 v34, 0x400000, v35
+; GFX11-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v34, 0x400000, v36
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
+; GFX11-NEXT:    v_bfe_u32 v37, v32, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v32
+; GFX11-NEXT:    v_bfe_u32 v50, v38, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
+; GFX11-NEXT:    v_bfe_u32 v34, v35, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v177
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_bfe_u32 v49, v37, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v50, v38
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v37
+; GFX11-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_bfe_u32 v48, v37, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v39, v38, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
+; GFX11-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v48, v37
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
+; GFX11-NEXT:    v_add_nc_u32_e32 v39, v39, v38
+; GFX11-NEXT:    v_lshl_or_b32 v177, v35, 16, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_or_b32_e32 v35, 0x400000, v37
+; GFX11-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v37, v33, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v39, v34, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v37, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v39, v34
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, v38, v35
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v33, v36, v49 :: v_dual_lshlrev_b32 v48, 16, v183
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v36, v38, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v38
+; GFX11-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v36, v38
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_or_b32_e32 v55, 0x400000, v48
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_bfe_u32 v50, v37, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v38, v39, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v54, 0x400000, v39
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
+; GFX11-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
+; GFX11-NEXT:    v_bfe_u32 v51, v48, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, v38, v39
+; GFX11-NEXT:    v_or_b32_e32 v53, 0x400000, v37
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
+; GFX11-NEXT:    v_bfe_u32 v52, v50, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v51, v51, v48
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v52, v52, v50
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
+; GFX11-NEXT:    v_or_b32_e32 v52, 0x400000, v50
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
+; GFX11-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
+; GFX11-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
+; GFX11-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
+; GFX11-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
+; GFX11-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
+; GFX11-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
+; GFX11-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
+; GFX11-NEXT:    v_lshl_or_b32 v183, v39, 16, v48
+; GFX11-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
+; GFX11-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
+; GFX11-NEXT:  .LBB43_3: ; %end
+; GFX11-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
+; GFX11-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
+; GFX11-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
+; GFX11-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
+; GFX11-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
+; GFX11-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
+; GFX11-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v19, v175
+; GFX11-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
+; GFX11-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
+; GFX11-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_load_b32 v184, off, s32
+; GFX11-NEXT:    scratch_load_b32 v175, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v174, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v173, off, s32 offset:12
+; GFX11-NEXT:    scratch_load_b32 v172, off, s32 offset:16
+; GFX11-NEXT:    scratch_load_b32 v171, off, s32 offset:20
+; GFX11-NEXT:    scratch_load_b32 v170, off, s32 offset:24
+; GFX11-NEXT:    scratch_load_b32 v169, off, s32 offset:28
+; GFX11-NEXT:    scratch_load_b32 v168, off, s32 offset:32
+; GFX11-NEXT:    scratch_load_b32 v159, off, s32 offset:36
+; GFX11-NEXT:    scratch_load_b32 v158, off, s32 offset:40
+; GFX11-NEXT:    scratch_load_b32 v157, off, s32 offset:44
+; GFX11-NEXT:    scratch_load_b32 v156, off, s32 offset:48
+; GFX11-NEXT:    scratch_load_b32 v155, off, s32 offset:52
+; GFX11-NEXT:    scratch_load_b32 v154, off, s32 offset:56
+; GFX11-NEXT:    scratch_load_b32 v153, off, s32 offset:60
+; GFX11-NEXT:    scratch_load_b32 v152, off, s32 offset:64
+; GFX11-NEXT:    scratch_load_b32 v143, off, s32 offset:68
+; GFX11-NEXT:    scratch_load_b32 v142, off, s32 offset:72
+; GFX11-NEXT:    scratch_load_b32 v141, off, s32 offset:76
+; GFX11-NEXT:    scratch_load_b32 v140, off, s32 offset:80
+; GFX11-NEXT:    scratch_load_b32 v139, off, s32 offset:84
+; GFX11-NEXT:    scratch_load_b32 v138, off, s32 offset:88
+; GFX11-NEXT:    scratch_load_b32 v137, off, s32 offset:92
+; GFX11-NEXT:    scratch_load_b32 v136, off, s32 offset:96
+; GFX11-NEXT:    scratch_load_b32 v127, off, s32 offset:100
+; GFX11-NEXT:    scratch_load_b32 v126, off, s32 offset:104
+; GFX11-NEXT:    scratch_load_b32 v125, off, s32 offset:108
+; GFX11-NEXT:    scratch_load_b32 v124, off, s32 offset:112
+; GFX11-NEXT:    scratch_load_b32 v123, off, s32 offset:116
+; GFX11-NEXT:    scratch_load_b32 v122, off, s32 offset:120
+; GFX11-NEXT:    scratch_load_b32 v121, off, s32 offset:124
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_load_b32 v120, off, s32 offset:128
+; GFX11-NEXT:    scratch_load_b32 v111, off, s32 offset:132
+; GFX11-NEXT:    scratch_load_b32 v110, off, s32 offset:136
+; GFX11-NEXT:    scratch_load_b32 v109, off, s32 offset:140
+; GFX11-NEXT:    scratch_load_b32 v108, off, s32 offset:144
+; GFX11-NEXT:    scratch_load_b32 v107, off, s32 offset:148
+; GFX11-NEXT:    scratch_load_b32 v106, off, s32 offset:152
+; GFX11-NEXT:    scratch_load_b32 v105, off, s32 offset:156
+; GFX11-NEXT:    scratch_load_b32 v104, off, s32 offset:160
+; GFX11-NEXT:    scratch_load_b32 v95, off, s32 offset:164
+; GFX11-NEXT:    scratch_load_b32 v94, off, s32 offset:168
+; GFX11-NEXT:    scratch_load_b32 v93, off, s32 offset:172
+; GFX11-NEXT:    scratch_load_b32 v92, off, s32 offset:176
+; GFX11-NEXT:    scratch_load_b32 v91, off, s32 offset:180
+; GFX11-NEXT:    scratch_load_b32 v90, off, s32 offset:184
+; GFX11-NEXT:    scratch_load_b32 v89, off, s32 offset:188
+; GFX11-NEXT:    scratch_load_b32 v88, off, s32 offset:192
+; GFX11-NEXT:    scratch_load_b32 v79, off, s32 offset:196
+; GFX11-NEXT:    scratch_load_b32 v78, off, s32 offset:200
+; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:204
+; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:208
+; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:212
+; GFX11-NEXT:    scratch_load_b32 v74, off, s32 offset:216
+; GFX11-NEXT:    scratch_load_b32 v73, off, s32 offset:220
+; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:224
+; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:228
+; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:232
+; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:236
+; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:240
+; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:244
+; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:248
+; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:252
+; GFX11-NEXT:    s_clause 0x8
+; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:256
+; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:260
+; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:264
+; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:268
+; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:272
+; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:276
+; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:280
+; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:284
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:288
+; GFX11-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
+; GFX11-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
+; GFX11-NEXT:    v_dual_mov_b32 v16, v183 :: v_dual_mov_b32 v21, v177
+; GFX11-NEXT:    v_dual_mov_b32 v24, v176 :: v_dual_mov_b32 v27, v181
+; GFX11-NEXT:    v_mov_b32_e32 v28, v182
+; GFX11-NEXT:    v_dual_mov_b32 v30, v179 :: v_dual_mov_b32 v31, v178
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB43_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
+; GFX11-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
+; GFX11-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
+; GFX11-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
+; GFX11-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
+; GFX11-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
+; GFX11-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
+; GFX11-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
+; GFX11-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
+; GFX11-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
+; GFX11-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
+; GFX11-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
+; GFX11-NEXT:    s_branch .LBB43_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -85886,1981 +82045,953 @@ define inreg <128 x i8> @bitcast_v16i64_to_v128i8_scalar(<16 x i64> inreg %a, i3
 ; GFX9-NEXT:    ; kill: killed $sgpr47
 ; GFX9-NEXT:    s_branch .LBB57_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v16i64_to_v128i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v16, s32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v17, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v18, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v19, s32 offset:12
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s30, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s96, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s97, 1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s98, 2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s99, 3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s100, 4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s101, 5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr18 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr19 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s102, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s103, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s104, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s50, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s51, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s52, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s53, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s54, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s55, 15
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s64, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s65, 17
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s66, 18
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s67, 19
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s68, 20
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s69, 21
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s70, 22
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s71, 23
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s80, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s81, 25
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s82, 26
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s83, 27
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s84, 28
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s85, 29
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s86, 30
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v16, s87, 31
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB57_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[4:5], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s100, s15, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s101, s40, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s29, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s29, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 31
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s4, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s28, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 30
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s4, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[6:7], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s27, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 29
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s27, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 28
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s26, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 27
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s25, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 26
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s6, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[10:11], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s25, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 25
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s6, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s51, s24, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s52, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s53, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s54, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 23
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s55, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s64, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s65, s22, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 22
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s66, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s67, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s68, s21, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 21
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s69, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s70, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s71, s19, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 20
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s80, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s81, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s82, s18, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 19
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s83, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s84, s17, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 18
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s85, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 17
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s86, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s87, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s96, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s97, s1, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 15
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s98, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s0, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 14
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[12:13], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[60:61], s[14:15], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 13
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[72:73], s[28:29], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[26:27], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[78:79], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[76:77], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[88:89], s[20:21], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 11
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s12, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s14, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s14, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s40, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s29, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s18, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s57
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB57_3
-; GFX11-TRUE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s40, s40, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s41, s41, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s14, s14, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s15, s15, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s12, s12, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s13, s13, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s10, s10, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s11, s11, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s8, s8, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s9, s9, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s7, s7, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s100, s15, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[4:5], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s101, s40, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s29, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s28, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 31
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s4, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s28, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s27, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 30
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s4, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[6:7], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 29
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s26, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 28
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s26, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 27
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s25, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 26
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s6, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[10:11], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s25, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 25
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s6, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s51, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s52, s24, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s53, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s54, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s55, s23, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 23
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s64, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s65, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s66, s21, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 22
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s67, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s68, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s69, s20, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 21
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s70, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s71, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s80, s19, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 20
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s81, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s82, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s17, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 19
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s83, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s84, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s85, s16, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 18
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s86, s3, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 17
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s87, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s96, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s97, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s98, s1, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 15
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s0, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s59, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 14
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[12:13], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[60:61], s[14:15], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 13
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[72:73], s[28:29], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[78:79], s[24:25], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[76:77], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[88:89], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 11
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s12, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s14, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s14, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s40, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s29, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s18, 8
-; GFX11-TRUE16-NEXT:  .LBB57_3: ; %end
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s56, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s99
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s30
-; GFX11-TRUE16-NEXT:    s_and_b32 s57, s57, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s56, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_or_b32 s56, s57, s56
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s47, s47, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s56, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s47
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s94
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s98
-; GFX11-TRUE16-NEXT:    s_and_b32 s46, s46, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s47, s47, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s56, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s46, s46, s47
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s97
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s96
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s46, s46, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s56, s56, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s57, s57, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s46
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s87
-; GFX11-TRUE16-NEXT:    s_or_b32 s56, s56, s57
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s46, s46, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s56, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s46
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s86
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s56
-; GFX11-TRUE16-NEXT:    s_and_b32 s46, s46, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s45, s45, 8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s0 :: v_dual_mov_b32 v4, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s45, s46, s45
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s44
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s45, s45, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s45
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s2 :: v_dual_mov_b32 v6, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s92
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s85
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s84
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s43
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s83
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s82
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s18
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s42
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s71
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s90
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s81
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s16, s16, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s0 :: v_dual_mov_b32 v8, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s70
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s80
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s16, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s88
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s69
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s68
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s66
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s67
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s64
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s22
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s65
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s55
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s76
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s53
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s23
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s18, s18, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s54
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s17, s17, 0xff
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s17, s17, s18
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s16, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s16, s17, 16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[3:6], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s52
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s51
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s78
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s50
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s49
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s48
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s39
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s38
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s74
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s36
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s35
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s18, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s19, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-TRUE16-NEXT:    s_or_b32 s17, s18, s19
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s104
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s72
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s17, v18, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s103
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s102
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s16, v18, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s101
-; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s62
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s17, v18, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s18, v18, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s19, v18, 4
-; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s41
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s18, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s19, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-TRUE16-NEXT:    s_or_b32 s17, s18, s19
-; GFX11-TRUE16-NEXT:    s_and_b32 s16, s16, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s1, v18, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s2, v18, 6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s60
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s14, v18, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s100
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, s14, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s15, s15, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s14, s15
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s12, v18, 10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s58
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s14, s14, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, s12, 0xff
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s12, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s13, v18, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s14, v18, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s15, v18, 13
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, s12, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s13, s13, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, s14, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s15, s15, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s12, s12, s13
-; GFX11-TRUE16-NEXT:    s_or_b32 s13, s14, s15
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, s12, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s13, s13, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s12, s13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s12, v19, 2
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s1, v18, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s2, v18, 15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s12
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 16
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s10, v18, 17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s11, v18, 18
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s10, s11
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 19
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s8, v18, 20
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s10, v19, 4
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s11, v19, 5
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s10, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s8, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s9, v18, 21
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s10, v18, 22
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s11, v18, 23
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s8, s9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s8, v19, 6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s1, v18, 24
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s2, v18, 25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 26
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s6, v18, 27
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s7, v18, 28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s6, s7
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s3, v18, 29
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s4, v18, 30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s6, v19, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s7, v19, 9
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s5, v18, 31
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s6, v19, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s7, v19, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s13, v19, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s9, v19, 7
-; GFX11-TRUE16-NEXT:    s_clause 0x2
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:112
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s104, v17, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s103, v17, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s102, v17, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s101, v17, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s100, v17, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s99, v17, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s98, v17, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s97, v17, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s96, v17, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s87, v16, 31
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s86, v16, 30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s85, v16, 29
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s84, v16, 28
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s83, v16, 27
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s82, v16, 26
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s81, v16, 25
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s80, v16, 24
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s71, v16, 23
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s70, v16, 22
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s69, v16, 21
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s68, v16, 20
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v16, 19
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v16, 18
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v16, 17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v16, 16
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v16, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v16, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v16, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v16, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v16, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v16, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v16, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v16, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v16, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v16, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v16, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v16, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v16, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v16, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v16, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v16, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v16, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v17, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v18, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v19, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB57_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr101_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr99_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr98_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr97_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr96_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr87_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr86_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr85_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr84_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr104_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr103_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr102_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 2
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 8
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 9
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 10
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 11
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 12
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 13
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 14
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 15
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 17
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 18
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 19
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 20
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 21
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 22
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 23
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 24
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 25
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 26
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 27
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 28
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 29
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 30
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v18, s58, 31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s58, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s100, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s101, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s100, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s101, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s100, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s101, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s100, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v19, s101, 9
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB57_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v16i64_to_v128i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v16, s32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v17, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v18, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v19, s32 offset:12
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s96, 0
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s97, 1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s98, 2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s99, 3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s100, 4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s101, 5
-; GFX11-FAKE16-NEXT:    s_mov_b32 s101, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr19 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr18 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s102, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s103, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s104, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s49, 9
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s50, 10
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s51, 11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s52, 12
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s53, 13
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s54, 14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s55, 15
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s64, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s65, 17
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s66, 18
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s67, 19
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s68, 20
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s69, 21
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s70, 22
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s71, 23
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s80, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s81, 25
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s82, 26
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s83, 27
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s84, 28
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s85, 29
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s86, 30
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v16, s87, 31
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB57_2
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[4:5], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s5, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s5, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 17
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s7, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 18
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s7, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s69, s6, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 19
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s6, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s9, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s9, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 20
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s9, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s8, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 21
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s11, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s50, s11, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 22
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s51, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s52, s10, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s53, s13, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 23
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s54, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s55, s13, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s64, s12, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s65, s12, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s66, s15, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s67, s15, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 25
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s68, s15, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s14, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 26
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s70, s41, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s71, s41, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s41, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 27
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s80, s40, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s40, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s81, s29, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 28
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s82, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s83, s29, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s84, s28, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 29
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s85, s28, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s86, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s27, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 30
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s87, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s96, s26, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s43, 31
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s97, s25, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 14
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s2, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s17, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 15
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[6:7], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s98, s1, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s17, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s99, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s1, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s17, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 13
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[8:9], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 10
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[76:77], s[26:27], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 11
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[10:11], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[88:89], s[24:25], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s43, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[14:15], 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[28:29], 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_branch .LBB57_3
-; GFX11-FAKE16-NEXT:  .LBB57_2:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    s_mov_b32 s101, -1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 0
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 1
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 2
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 3
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr100
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr99
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr98
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr97
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr96
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr87
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr86
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr85
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr103
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr102
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr104
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 10
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 12
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_lo
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_lo, 14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, vcc_hi, 15
-; GFX11-FAKE16-NEXT:  .LBB57_3: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s101
-; GFX11-FAKE16-NEXT:    s_mov_b32 s101, s104
-; GFX11-FAKE16-NEXT:    s_mov_b32 s104, s57
-; GFX11-FAKE16-NEXT:    s_mov_b32 s57, s69
-; GFX11-FAKE16-NEXT:    s_mov_b32 s69, s42
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB57_5
-; GFX11-FAKE16-NEXT:  ; %bb.4: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s40, s40, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s41, s41, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s14, s14, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s15, s15, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s12, s12, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s13, s13, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s11, s11, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s9, s9, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s7, s7, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[4:5], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s101, s5, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s5, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 17
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s7, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 18
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s7, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s6, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 19
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s6, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s9, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s9, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 20
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s9, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s8, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 21
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s11, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s50, s11, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 22
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s51, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s52, s10, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s53, s13, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 23
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s54, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s55, s13, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s64, s12, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s65, s12, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s66, s15, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s67, s15, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 25
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s68, s15, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s14, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 26
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s70, s41, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s71, s41, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s41, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 27
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s80, s40, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s40, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s81, s29, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 28
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s82, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s83, s29, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s84, s28, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 29
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s85, s28, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s86, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s27, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 30
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s87, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s96, s26, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s42, 31
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s97, s25, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 14
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s69, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s2, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 15
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[6:7], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s98, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s99, s1, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 13
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[8:9], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s1, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 10
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 11
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[10:11], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[76:77], s[26:27], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[88:89], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v18, s42, 9
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[14:15], 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[28:29], 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s62, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v19, s63, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
-; GFX11-FAKE16-NEXT:  .LBB57_5: ; %end
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s43, s43, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, s74, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s43
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s43, s94, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s42, s42, s43
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s45, s45, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s42, s42, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s42
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s42, v18, 9
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s45, s30, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s44, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s44, s44, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s42, s42, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s44, s44, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s42
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s42, v18, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s43, v18, 7
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s44
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s44, s100, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s45, s98, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s44
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s99, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, s42, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s44, s44, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s43, s43, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s44, s44, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s42, s42, s43
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s44
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s42, s42, 16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v18, 6
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s42
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v18, 5
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s3, v18, 2
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s92, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v18, 4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v18, 0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v18, 3
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s17, v19, 29
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s100, v17, 4
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s99, v17, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v18, 1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v19, 28
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s18, v19, 19
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s90, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v19, 31
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s19, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s16, s16, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v19, 30
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s78, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s18, s18, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s19, s86, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s17, v19, 21
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s16, s16, 16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s98, v17, 2
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v19, 27
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s3, v19, 24
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v19, 22
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s17, 8
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v19, 26
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s21, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s86, v16, 30
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v19, 25
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v16, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v16, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v19, 23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s1, v19, 18
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s62, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s23, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s17, v19, 20
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s16, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s17, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s17, s17, s18
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s16, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s97, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s2, v19, 17
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s88, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s69, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s18, s72, 0xff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s97, v17, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s69, v16, 21
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s3, v19, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s25, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s73, 0xff
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s96, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s76, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s27, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s87, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-FAKE16-NEXT:    s_or_b32 s17, s18, s19
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s16, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s17, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v19, 0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s85, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s84, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s16, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s17, v19, 1
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s29, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s83, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s82, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s81, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s18, v19, 2
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s40, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s61, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s80, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s18, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s19, v19, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s41, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s60, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s18, s71, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s19, s70, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s16, s16, s17
-; GFX11-FAKE16-NEXT:    s_or_b32 s17, s18, s19
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s16, s16, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s17, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s16, s17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s16, v19, 4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s14, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s58, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s59, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s16, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s15, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s68, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, s67, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s15, s66, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s14, s15
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s14, v19, 6
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s12, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s65, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, s64, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s14, s14, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s15, v19, 7
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s12, s14
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, s13, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s13, s55, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, s54, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s15, s53, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s12, s12, s13
-; GFX11-FAKE16-NEXT:    s_or_b32 s13, s14, s15
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, s12, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s13, s13, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s12, s13
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s12, v19, 8
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s10, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s52, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s51, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s12, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s11, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s50, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s49, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s48, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s10, s11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s10, v19, 10
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s8, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s39, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s38, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s10, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s11, v19, 11
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s8, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s9, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s37, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s36, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s35, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s8, s9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s8, v19, 12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s6, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s56, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s57, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s8, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s7, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s34, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_hi, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s46, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s6, s7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s6, v19, 14
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s4, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s47, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s104, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s7, v19, 15
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s5, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s103, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s102, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s101, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:64
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s17, v19, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s13, v19, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s9, v19, 13
-; GFX11-FAKE16-NEXT:    s_clause 0x2
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:112
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s104, v17, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s103, v17, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s102, v17, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s101, v17, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s96, v17, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s87, v16, 31
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s85, v16, 29
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s84, v16, 28
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s83, v16, 27
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s82, v16, 26
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s81, v16, 25
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s80, v16, 24
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s71, v16, 23
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s70, v16, 22
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s68, v16, 20
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v16, 19
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v16, 18
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v16, 17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v16, 16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v16, 15
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v16, 14
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v16, 13
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v16, 12
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v16, 11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v16, 10
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v16, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v16, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v16, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v16, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v16, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v16, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v16, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v16, 2
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v16, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v17, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v18, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v19, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v16i64_to_v128i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b32 off, v16, s32
+; GFX11-NEXT:    scratch_store_b32 off, v17, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v18, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v19, s32 offset:12
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v16, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v17, s96, 0
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
+; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s41, v2
+; GFX11-NEXT:    v_writelane_b32 v16, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v17, s97, 1
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s15, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
+; GFX11-NEXT:    v_writelane_b32 v16, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v17, s98, 2
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v8
+; GFX11-NEXT:    v_writelane_b32 v16, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v17, s99, 3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
+; GFX11-NEXT:    v_writelane_b32 v16, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v17, s100, 4
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v14
+; GFX11-NEXT:    v_writelane_b32 v16, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v17, s101, 5
+; GFX11-NEXT:    s_mov_b32 s101, 0
+; GFX11-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
+; GFX11-NEXT:    ; implicit-def: $vgpr19 : SGPR spill to VGPR lane
+; GFX11-NEXT:    ; implicit-def: $vgpr18 : SGPR spill to VGPR lane
+; GFX11-NEXT:    v_writelane_b32 v16, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v17, s102, 6
+; GFX11-NEXT:    v_writelane_b32 v16, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v17, s103, 7
+; GFX11-NEXT:    v_writelane_b32 v16, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v17, s104, 8
+; GFX11-NEXT:    v_writelane_b32 v16, s49, 9
+; GFX11-NEXT:    v_writelane_b32 v16, s50, 10
+; GFX11-NEXT:    v_writelane_b32 v16, s51, 11
+; GFX11-NEXT:    v_writelane_b32 v16, s52, 12
+; GFX11-NEXT:    v_writelane_b32 v16, s53, 13
+; GFX11-NEXT:    v_writelane_b32 v16, s54, 14
+; GFX11-NEXT:    v_writelane_b32 v16, s55, 15
+; GFX11-NEXT:    v_writelane_b32 v16, s64, 16
+; GFX11-NEXT:    v_writelane_b32 v16, s65, 17
+; GFX11-NEXT:    v_writelane_b32 v16, s66, 18
+; GFX11-NEXT:    v_writelane_b32 v16, s67, 19
+; GFX11-NEXT:    v_writelane_b32 v16, s68, 20
+; GFX11-NEXT:    v_writelane_b32 v16, s69, 21
+; GFX11-NEXT:    v_writelane_b32 v16, s70, 22
+; GFX11-NEXT:    v_writelane_b32 v16, s71, 23
+; GFX11-NEXT:    v_writelane_b32 v16, s80, 24
+; GFX11-NEXT:    v_writelane_b32 v16, s81, 25
+; GFX11-NEXT:    v_writelane_b32 v16, s82, 26
+; GFX11-NEXT:    v_writelane_b32 v16, s83, 27
+; GFX11-NEXT:    v_writelane_b32 v16, s84, 28
+; GFX11-NEXT:    v_writelane_b32 v16, s85, 29
+; GFX11-NEXT:    v_writelane_b32 v16, s86, 30
+; GFX11-NEXT:    v_writelane_b32 v16, s87, 31
+; GFX11-NEXT:    s_cbranch_scc0 .LBB57_2
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s43, s25, 8
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[4:5], 24
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s104, s5, 24
+; GFX11-NEXT:    s_lshr_b32 s102, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s103, s5, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 17
+; GFX11-NEXT:    s_lshr_b32 s43, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s57, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 8
+; GFX11-NEXT:    s_lshr_b32 s46, s7, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 18
+; GFX11-NEXT:    s_lshr_b32 s43, s23, 24
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s34, s7, 8
+; GFX11-NEXT:    s_lshr_b32 s69, s6, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 19
+; GFX11-NEXT:    s_lshr_b32 s43, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s6, 8
+; GFX11-NEXT:    s_lshr_b32 s35, s9, 24
+; GFX11-NEXT:    s_lshr_b32 s36, s9, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 20
+; GFX11-NEXT:    s_lshr_b32 s43, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s37, s9, 8
+; GFX11-NEXT:    s_lshr_b32 s38, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s39, s8, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 21
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s11, 24
+; GFX11-NEXT:    s_lshr_b32 s49, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s50, s11, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 22
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s51, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s52, s10, 8
+; GFX11-NEXT:    s_lshr_b32 s53, s13, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 23
+; GFX11-NEXT:    s_lshr_b32 s43, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s54, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s55, s13, 8
+; GFX11-NEXT:    s_lshr_b32 s64, s12, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s65, s12, 8
+; GFX11-NEXT:    s_lshr_b32 s66, s15, 24
+; GFX11-NEXT:    s_lshr_b32 s67, s15, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 25
+; GFX11-NEXT:    s_lshr_b32 s43, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s68, s15, 8
+; GFX11-NEXT:    s_lshr_b32 s59, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s14, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 26
+; GFX11-NEXT:    s_lshr_b32 s43, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s70, s41, 24
+; GFX11-NEXT:    s_lshr_b32 s71, s41, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s41, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 27
+; GFX11-NEXT:    s_lshr_b32 s43, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s80, s40, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s40, 8
+; GFX11-NEXT:    s_lshr_b32 s81, s29, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 28
+; GFX11-NEXT:    s_lshr_b32 s43, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s82, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s83, s29, 8
+; GFX11-NEXT:    s_lshr_b32 s84, s28, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 29
+; GFX11-NEXT:    s_lshr_b32 s43, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s85, s28, 8
+; GFX11-NEXT:    s_lshr_b32 s86, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s72, s27, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 30
+; GFX11-NEXT:    s_lshr_b32 s43, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s87, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s73, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s96, s26, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s43, 31
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s97, s25, 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 0
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 14
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s2, 16
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 1
+; GFX11-NEXT:    s_lshr_b32 s43, s17, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 15
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[6:7], 24
+; GFX11-NEXT:    s_lshr_b32 s98, s1, 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 2
+; GFX11-NEXT:    s_lshr_b32 s43, s17, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 12
+; GFX11-NEXT:    s_lshr_b32 s99, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s100, s1, 8
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 3
+; GFX11-NEXT:    s_lshr_b32 s43, s17, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 13
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[8:9], 24
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 4
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 10
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[76:77], s[26:27], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 5
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 11
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[10:11], 24
+; GFX11-NEXT:    s_lshr_b64 s[88:89], s[24:25], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 6
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 8
+; GFX11-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 7
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 9
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
+; GFX11-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 6
+; GFX11-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s43, 9
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 7
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[14:15], 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 4
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 5
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 2
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 3
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[28:29], 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 0
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 1
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
+; GFX11-NEXT:    s_branch .LBB57_3
+; GFX11-NEXT:  .LBB57_2:
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    s_mov_b32 s101, -1
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 0
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 1
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 2
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 3
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 4
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 5
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $sgpr100
+; GFX11-NEXT:    ; implicit-def: $sgpr99
+; GFX11-NEXT:    ; implicit-def: $sgpr98
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr97
+; GFX11-NEXT:    ; implicit-def: $sgpr96
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr87
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr86
+; GFX11-NEXT:    ; implicit-def: $sgpr85
+; GFX11-NEXT:    ; implicit-def: $sgpr84
+; GFX11-NEXT:    ; implicit-def: $sgpr83
+; GFX11-NEXT:    ; implicit-def: $sgpr82
+; GFX11-NEXT:    ; implicit-def: $sgpr81
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr80
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr71
+; GFX11-NEXT:    ; implicit-def: $sgpr70
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr68
+; GFX11-NEXT:    ; implicit-def: $sgpr67
+; GFX11-NEXT:    ; implicit-def: $sgpr66
+; GFX11-NEXT:    ; implicit-def: $sgpr65
+; GFX11-NEXT:    ; implicit-def: $sgpr64
+; GFX11-NEXT:    ; implicit-def: $sgpr55
+; GFX11-NEXT:    ; implicit-def: $sgpr54
+; GFX11-NEXT:    ; implicit-def: $sgpr53
+; GFX11-NEXT:    ; implicit-def: $sgpr52
+; GFX11-NEXT:    ; implicit-def: $sgpr51
+; GFX11-NEXT:    ; implicit-def: $sgpr50
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr69
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr103
+; GFX11-NEXT:    ; implicit-def: $sgpr102
+; GFX11-NEXT:    ; implicit-def: $sgpr104
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 6
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 7
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 8
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 9
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 10
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 11
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 12
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 13
+; GFX11-NEXT:    ; implicit-def: $vcc_lo
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_lo, 14
+; GFX11-NEXT:    v_writelane_b32 v19, vcc_hi, 15
+; GFX11-NEXT:  .LBB57_3: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s101
+; GFX11-NEXT:    s_mov_b32 s101, s104
+; GFX11-NEXT:    s_mov_b32 s104, s57
+; GFX11-NEXT:    s_mov_b32 s57, s69
+; GFX11-NEXT:    s_mov_b32 s69, s42
+; GFX11-NEXT:    s_cbranch_vccnz .LBB57_5
+; GFX11-NEXT:  ; %bb.4: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s40, s40, 3
+; GFX11-NEXT:    s_addc_u32 s41, s41, 0
+; GFX11-NEXT:    s_add_u32 s14, s14, 3
+; GFX11-NEXT:    s_addc_u32 s15, s15, 0
+; GFX11-NEXT:    s_add_u32 s12, s12, 3
+; GFX11-NEXT:    s_addc_u32 s13, s13, 0
+; GFX11-NEXT:    s_add_u32 s10, s10, 3
+; GFX11-NEXT:    s_addc_u32 s11, s11, 0
+; GFX11-NEXT:    s_add_u32 s8, s8, 3
+; GFX11-NEXT:    s_addc_u32 s9, s9, 0
+; GFX11-NEXT:    s_add_u32 s6, s6, 3
+; GFX11-NEXT:    s_addc_u32 s7, s7, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 8
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[4:5], 24
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s101, s5, 24
+; GFX11-NEXT:    s_lshr_b32 s102, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s103, s5, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 17
+; GFX11-NEXT:    s_lshr_b32 s42, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s104, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 8
+; GFX11-NEXT:    s_lshr_b32 s46, s7, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 18
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 24
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s34, s7, 8
+; GFX11-NEXT:    s_lshr_b32 s57, s6, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 19
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s6, 8
+; GFX11-NEXT:    s_lshr_b32 s35, s9, 24
+; GFX11-NEXT:    s_lshr_b32 s36, s9, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 20
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s37, s9, 8
+; GFX11-NEXT:    s_lshr_b32 s38, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s39, s8, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 21
+; GFX11-NEXT:    s_lshr_b32 s42, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s11, 24
+; GFX11-NEXT:    s_lshr_b32 s49, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s50, s11, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 22
+; GFX11-NEXT:    s_lshr_b32 s42, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s51, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s52, s10, 8
+; GFX11-NEXT:    s_lshr_b32 s53, s13, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 23
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s54, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s55, s13, 8
+; GFX11-NEXT:    s_lshr_b32 s64, s12, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s65, s12, 8
+; GFX11-NEXT:    s_lshr_b32 s66, s15, 24
+; GFX11-NEXT:    s_lshr_b32 s67, s15, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 25
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s68, s15, 8
+; GFX11-NEXT:    s_lshr_b32 s59, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s14, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 26
+; GFX11-NEXT:    s_lshr_b32 s42, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s70, s41, 24
+; GFX11-NEXT:    s_lshr_b32 s71, s41, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s41, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 27
+; GFX11-NEXT:    s_lshr_b32 s42, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s80, s40, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s40, 8
+; GFX11-NEXT:    s_lshr_b32 s81, s29, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 28
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s82, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s83, s29, 8
+; GFX11-NEXT:    s_lshr_b32 s84, s28, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 29
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s85, s28, 8
+; GFX11-NEXT:    s_lshr_b32 s86, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s72, s27, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 30
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s87, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s73, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s96, s26, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s42, 31
+; GFX11-NEXT:    s_lshr_b32 s42, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s97, s25, 24
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 0
+; GFX11-NEXT:    s_lshr_b32 s42, s18, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 14
+; GFX11-NEXT:    s_lshr_b32 s69, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s2, 16
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 1
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 15
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[6:7], 24
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 8
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 2
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 12
+; GFX11-NEXT:    s_lshr_b32 s98, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s99, s1, 16
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 13
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[8:9], 24
+; GFX11-NEXT:    s_lshr_b32 s100, s1, 8
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 4
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 10
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 8
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 5
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 11
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[10:11], 24
+; GFX11-NEXT:    s_lshr_b64 s[76:77], s[26:27], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 6
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 24
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 8
+; GFX11-NEXT:    s_lshr_b64 s[88:89], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 7
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 9
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
+; GFX11-NEXT:    s_lshr_b64 s[90:91], s[18:19], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 8
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 6
+; GFX11-NEXT:    s_lshr_b64 s[92:93], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[94:95], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
+; GFX11-NEXT:    v_writelane_b32 v18, s42, 9
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 7
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[14:15], 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 4
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 5
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 2
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 3
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[28:29], 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v19, s62, 0
+; GFX11-NEXT:    v_writelane_b32 v19, s63, 1
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
+; GFX11-NEXT:  .LBB57_5: ; %end
+; GFX11-NEXT:    s_lshl_b32 s43, s43, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_and_b32 s42, s74, 0xff
+; GFX11-NEXT:    s_or_b32 s2, s2, s43
+; GFX11-NEXT:    s_lshl_b32 s43, s94, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_or_b32 s42, s42, s43
+; GFX11-NEXT:    s_lshl_b32 s45, s45, 8
+; GFX11-NEXT:    s_lshl_b32 s42, s42, 16
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_or_b32 s2, s2, s42
+; GFX11-NEXT:    v_readlane_b32 s42, v18, 9
+; GFX11-NEXT:    s_or_b32 s0, s0, s45
+; GFX11-NEXT:    s_lshl_b32 s45, s30, 8
+; GFX11-NEXT:    s_and_b32 s44, s44, 0xff
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xff
+; GFX11-NEXT:    s_or_b32 s44, s44, s45
+; GFX11-NEXT:    s_lshl_b32 s42, s42, 8
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s44, s44, 16
+; GFX11-NEXT:    s_or_b32 s3, s3, s42
+; GFX11-NEXT:    v_readlane_b32 s42, v18, 8
+; GFX11-NEXT:    v_readlane_b32 s43, v18, 7
+; GFX11-NEXT:    s_or_b32 s0, s0, s44
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xff
+; GFX11-NEXT:    s_lshl_b32 s44, s100, 8
+; GFX11-NEXT:    s_lshl_b32 s45, s98, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s44
+; GFX11-NEXT:    s_and_b32 s44, s99, 0xff
+; GFX11-NEXT:    s_and_b32 s42, s42, 0xff
+; GFX11-NEXT:    s_or_b32 s44, s44, s45
+; GFX11-NEXT:    s_lshl_b32 s43, s43, 8
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s44, s44, 16
+; GFX11-NEXT:    s_or_b32 s42, s42, s43
+; GFX11-NEXT:    s_or_b32 s1, s1, s44
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s42, s42, 16
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_readlane_b32 s0, v18, 6
+; GFX11-NEXT:    s_or_b32 s3, s3, s42
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    v_readlane_b32 s2, v18, 5
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 8
+; GFX11-NEXT:    s_and_b32 s1, s16, 0xff
+; GFX11-NEXT:    v_readlane_b32 s3, v18, 2
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_lshl_b32 s1, s92, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    v_readlane_b32 s2, v18, 4
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s17, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 8
+; GFX11-NEXT:    v_readlane_b32 s16, v18, 0
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_readlane_b32 s2, v18, 3
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    v_readlane_b32 s17, v19, 29
+; GFX11-NEXT:    s_and_b32 s16, s16, 0xff
+; GFX11-NEXT:    v_readlane_b32 s100, v17, 4
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    v_readlane_b32 s99, v17, 3
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_and_b32 s3, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_lshl_b32 s17, s17, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_readlane_b32 s2, v18, 1
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_readlane_b32 s0, v19, 28
+; GFX11-NEXT:    s_and_b32 s1, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 8
+; GFX11-NEXT:    v_readlane_b32 s18, v19, 19
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_lshl_b32 s3, s90, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_or_b32 s3, s16, s3
+; GFX11-NEXT:    v_readlane_b32 s16, v19, 31
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_and_b32 s3, s19, 0xff
+; GFX11-NEXT:    s_lshl_b32 s16, s16, 8
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s3, s3, s16
+; GFX11-NEXT:    v_readlane_b32 s16, v19, 30
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s78, 8
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s18, s18, 8
+; GFX11-NEXT:    s_and_b32 s16, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s19, s86, 8
+; GFX11-NEXT:    s_or_b32 s16, s16, s17
+; GFX11-NEXT:    v_readlane_b32 s17, v19, 21
+; GFX11-NEXT:    s_lshl_b32 s16, s16, 16
+; GFX11-NEXT:    v_readlane_b32 s98, v17, 2
+; GFX11-NEXT:    s_or_b32 s3, s3, s16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    v_readlane_b32 s2, v19, 27
+; GFX11-NEXT:    v_readlane_b32 s3, v19, 24
+; GFX11-NEXT:    v_readlane_b32 s16, v19, 22
+; GFX11-NEXT:    s_lshl_b32 s17, s17, 8
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    v_readlane_b32 s2, v19, 26
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s16, s16, 0xff
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s21, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 8
+; GFX11-NEXT:    v_readlane_b32 s86, v16, 30
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_readlane_b32 s2, v19, 25
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    v_readlane_b32 s31, v16, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v16, 0
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_and_b32 s3, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_readlane_b32 s2, v19, 23
+; GFX11-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
+; GFX11-NEXT:    v_readlane_b32 s1, v19, 18
+; GFX11-NEXT:    s_and_b32 s0, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_lshl_b32 s3, s62, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_or_b32 s3, s16, s3
+; GFX11-NEXT:    s_and_b32 s16, s23, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s16, s16, s17
+; GFX11-NEXT:    v_readlane_b32 s17, v19, 20
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_and_b32 s3, s16, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s17, s17, 0xff
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s17, s17, s18
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s16, s17, 16
+; GFX11-NEXT:    s_lshl_b32 s17, s97, 8
+; GFX11-NEXT:    s_or_b32 s3, s3, s16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
+; GFX11-NEXT:    v_readlane_b32 s2, v19, 17
+; GFX11-NEXT:    s_lshl_b32 s3, s88, 8
+; GFX11-NEXT:    s_and_b32 s16, s69, 0xff
+; GFX11-NEXT:    s_and_b32 s18, s72, 0xff
+; GFX11-NEXT:    v_readlane_b32 s97, v17, 1
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    v_readlane_b32 s69, v16, 21
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    v_readlane_b32 s3, v19, 16
+; GFX11-NEXT:    s_and_b32 s2, s25, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s16, s73, 0xff
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s96, 8
+; GFX11-NEXT:    s_lshl_b32 s17, s76, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    s_and_b32 s16, s27, 0xff
+; GFX11-NEXT:    s_lshl_b32 s17, s87, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_or_b32 s16, s16, s17
+; GFX11-NEXT:    s_or_b32 s17, s18, s19
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s16, s16, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s17, s17, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    v_readlane_b32 s16, v19, 0
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    s_and_b32 s0, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s85, 8
+; GFX11-NEXT:    s_and_b32 s2, s84, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s16, 8
+; GFX11-NEXT:    v_readlane_b32 s17, v19, 1
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s29, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s83, 8
+; GFX11-NEXT:    s_and_b32 s16, s82, 0xff
+; GFX11-NEXT:    s_lshl_b32 s17, s81, 8
+; GFX11-NEXT:    v_readlane_b32 s18, v19, 2
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s40, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s61, 8
+; GFX11-NEXT:    s_and_b32 s16, s80, 0xff
+; GFX11-NEXT:    s_lshl_b32 s17, s18, 8
+; GFX11-NEXT:    v_readlane_b32 s19, v19, 3
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    s_and_b32 s16, s41, 0xff
+; GFX11-NEXT:    s_lshl_b32 s17, s60, 8
+; GFX11-NEXT:    s_and_b32 s18, s71, 0xff
+; GFX11-NEXT:    s_lshl_b32 s19, s70, 8
+; GFX11-NEXT:    s_or_b32 s16, s16, s17
+; GFX11-NEXT:    s_or_b32 s17, s18, s19
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s16, s16, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s17, s17, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s16, s17
+; GFX11-NEXT:    v_readlane_b32 s16, v19, 4
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    s_and_b32 s0, s14, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s58, 8
+; GFX11-NEXT:    s_and_b32 s2, s59, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s16, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s15, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s68, 8
+; GFX11-NEXT:    s_and_b32 s14, s67, 0xff
+; GFX11-NEXT:    s_lshl_b32 s15, s66, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s14, s15
+; GFX11-NEXT:    v_readlane_b32 s14, v19, 6
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s12, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s65, 8
+; GFX11-NEXT:    s_and_b32 s12, s64, 0xff
+; GFX11-NEXT:    s_lshl_b32 s14, s14, 8
+; GFX11-NEXT:    v_readlane_b32 s15, v19, 7
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s12, s14
+; GFX11-NEXT:    s_and_b32 s12, s13, 0xff
+; GFX11-NEXT:    s_lshl_b32 s13, s55, 8
+; GFX11-NEXT:    s_and_b32 s14, s54, 0xff
+; GFX11-NEXT:    s_lshl_b32 s15, s53, 8
+; GFX11-NEXT:    s_or_b32 s12, s12, s13
+; GFX11-NEXT:    s_or_b32 s13, s14, s15
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s12, s12, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s13, s13, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s12, s13
+; GFX11-NEXT:    v_readlane_b32 s12, v19, 8
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
+; GFX11-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
+; GFX11-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
+; GFX11-NEXT:    s_and_b32 s0, s10, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s52, 8
+; GFX11-NEXT:    s_and_b32 s2, s51, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s12, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s11, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s50, 8
+; GFX11-NEXT:    s_and_b32 s10, s49, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s48, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s10, s11
+; GFX11-NEXT:    v_readlane_b32 s10, v19, 10
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s8, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s39, 8
+; GFX11-NEXT:    s_and_b32 s8, s38, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s10, 8
+; GFX11-NEXT:    v_readlane_b32 s11, v19, 11
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s8, s10
+; GFX11-NEXT:    s_and_b32 s8, s9, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s37, 8
+; GFX11-NEXT:    s_and_b32 s10, s36, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s35, 8
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    s_or_b32 s9, s10, s11
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s8, s8, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s9, s9, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s8, s9
+; GFX11-NEXT:    v_readlane_b32 s8, v19, 12
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    s_and_b32 s0, s6, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s56, 8
+; GFX11-NEXT:    s_and_b32 s2, s57, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s8, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s7, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s34, 8
+; GFX11-NEXT:    s_and_b32 s6, vcc_hi, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s46, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s6, s7
+; GFX11-NEXT:    v_readlane_b32 s6, v19, 14
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s4, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s47, 8
+; GFX11-NEXT:    s_and_b32 s4, s104, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 8
+; GFX11-NEXT:    v_readlane_b32 s7, v19, 15
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s6
+; GFX11-NEXT:    s_and_b32 s4, s5, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s103, 8
+; GFX11-NEXT:    s_and_b32 s6, s102, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s101, 8
+; GFX11-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-NEXT:    s_or_b32 s5, s6, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s4, s4, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:64
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    v_readlane_b32 s17, v19, 5
+; GFX11-NEXT:    v_readlane_b32 s13, v19, 9
+; GFX11-NEXT:    v_readlane_b32 s9, v19, 13
+; GFX11-NEXT:    s_clause 0x2
+; GFX11-NEXT:    scratch_store_b128 v0, v[9:12], off offset:80
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:96
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:112
+; GFX11-NEXT:    v_readlane_b32 s104, v17, 8
+; GFX11-NEXT:    v_readlane_b32 s103, v17, 7
+; GFX11-NEXT:    v_readlane_b32 s102, v17, 6
+; GFX11-NEXT:    v_readlane_b32 s101, v17, 5
+; GFX11-NEXT:    v_readlane_b32 s96, v17, 0
+; GFX11-NEXT:    v_readlane_b32 s87, v16, 31
+; GFX11-NEXT:    v_readlane_b32 s85, v16, 29
+; GFX11-NEXT:    v_readlane_b32 s84, v16, 28
+; GFX11-NEXT:    v_readlane_b32 s83, v16, 27
+; GFX11-NEXT:    v_readlane_b32 s82, v16, 26
+; GFX11-NEXT:    v_readlane_b32 s81, v16, 25
+; GFX11-NEXT:    v_readlane_b32 s80, v16, 24
+; GFX11-NEXT:    v_readlane_b32 s71, v16, 23
+; GFX11-NEXT:    v_readlane_b32 s70, v16, 22
+; GFX11-NEXT:    v_readlane_b32 s68, v16, 20
+; GFX11-NEXT:    v_readlane_b32 s67, v16, 19
+; GFX11-NEXT:    v_readlane_b32 s66, v16, 18
+; GFX11-NEXT:    v_readlane_b32 s65, v16, 17
+; GFX11-NEXT:    v_readlane_b32 s64, v16, 16
+; GFX11-NEXT:    v_readlane_b32 s55, v16, 15
+; GFX11-NEXT:    v_readlane_b32 s54, v16, 14
+; GFX11-NEXT:    v_readlane_b32 s53, v16, 13
+; GFX11-NEXT:    v_readlane_b32 s52, v16, 12
+; GFX11-NEXT:    v_readlane_b32 s51, v16, 11
+; GFX11-NEXT:    v_readlane_b32 s50, v16, 10
+; GFX11-NEXT:    v_readlane_b32 s49, v16, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v16, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v16, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v16, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v16, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v16, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v16, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v16, 2
+; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_load_b32 v16, off, s32
+; GFX11-NEXT:    scratch_load_b32 v17, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v18, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v19, off, s32 offset:12
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -97102,462 +92233,482 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v16i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:436
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:320
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v86, v0
+; GFX11-TRUE16-NEXT:    s_clause 0x1f
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:476
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:472
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:468
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:464
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:460
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:456
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:452
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:448
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:444
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:352
+; GFX11-TRUE16-NEXT:    s_clause 0x7
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:320
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v26 :: v_dual_mov_b32 v35, v24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v22 :: v_dual_mov_b32 v37, v20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v18 :: v_dual_mov_b32 v39, v16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v14 :: v_dual_mov_b32 v49, v12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v10 :: v_dual_mov_b32 v51, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v6 :: v_dual_mov_b32 v53, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, v2 :: v_dual_mov_b32 v55, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
 ; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:316
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:312
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:308
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:304
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:300
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:296
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:292
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v2, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v16, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v18, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v20, off, s32 offset:72
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v22, off, s32 offset:80
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v24, off, s32 offset:88
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v26, off, s32 offset:96
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v28, off, s32 offset:104
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v30, off, s32 offset:112
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:120
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:128
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:136
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:144
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:152
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:160
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:168
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:176
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:184
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:200
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:208
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:216
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:224
 ; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v94, off, s32 offset:240
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v95, off, s32 offset:248
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v104, off, s32 offset:256
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v105, off, s32 offset:264
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v106, off, s32 offset:272
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v107, off, s32 offset:280
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v108, off, s32 offset:288
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v109, off, s32 offset:296
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v110, off, s32 offset:304
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v111, off, s32 offset:312
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:308
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:300
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:292
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:284
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:276
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:268
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:260
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:252
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:244
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:228
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:220
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:212
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:204
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:196
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:188
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:180
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:172
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:164
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:156
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:148
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:140
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:132
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v89, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v88, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v90, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v91, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v93, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v92, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v28 :: v_dual_mov_b32 v34, v27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v26 :: v_dual_mov_b32 v37, v25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v24 :: v_dual_mov_b32 v38, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v22 :: v_dual_mov_b32 v49, v21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v20 :: v_dual_mov_b32 v51, v19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v18 :: v_dual_mov_b32 v53, v17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v55, v16 :: v_dual_mov_b32 v54, v15
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, v14 :: v_dual_mov_b32 v65, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v67, v12 :: v_dual_mov_b32 v66, v11
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v68, v10 :: v_dual_mov_b32 v69, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v71, v8 :: v_dual_mov_b32 v70, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v80, v6 :: v_dual_mov_b32 v81, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v83, v4 :: v_dual_mov_b32 v82, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v84, v2 :: v_dual_mov_b32 v85, v1
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:124
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:116
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:108
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:100
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:92
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:84
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:76
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:68
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v89, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v90, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v91, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v92, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v93, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v76, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v77, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v78, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v79, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v88, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v63, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v72, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v73, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v74, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v75, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v57, 8, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v58, 8, v4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v59, 8, v6
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v60, 8, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v61, 8, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v40, 8, v12
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 8, v14
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v42, 8, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v43, 8, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v44, 8, v20
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v167, 8, v22
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v176, 8, v24
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v177, 8, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v178, 8, v28
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v179, 8, v30
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v149, 8, v31
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v150, 8, v87
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(61)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v151, 8, v96
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(60)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v160, 8, v97
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(59)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v161, 8, v98
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(58)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v132, 8, v99
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(57)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v133, 8, v113
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(56)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v134, 8, v114
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(55)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v135, 8, v115
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(54)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v144, 8, v116
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(53)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v119, 8, v117
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(52)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v128, 8, v128
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(51)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v129, 8, v129
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(50)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v130, 8, v130
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(49)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v131, 8, v131
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(48)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v113, 8, v94
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(47)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v114, 8, v95
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(46)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v115, 8, v104
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(45)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v116, 8, v105
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(44)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v117, 8, v106
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(43)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v107
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(42)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v108
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(41)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v109
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(40)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v98, 8, v110
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(39)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v99, 8, v111
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB59_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v71
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v53
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v90
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v91
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v68
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v65
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v67
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v53
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s26, 0xff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v49
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s27, 8
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v76
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v77
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v64
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v78
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v79
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v49
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v51
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v37
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v88
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v63
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v37
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v72
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v73
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v33
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v33
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v74
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v75
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v91
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v0, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v32
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v92
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v93
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v62
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v57
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v58
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v79
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v90
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v88
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v89
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v56
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v47
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v59
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v60
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v75
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v78
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v77
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v76
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v45
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v61
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v40
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v63
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v74
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v72
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v73
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v183
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v182
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v41
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v42
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v59
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v62
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v60
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v61
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v43
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v44
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v47
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v56
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v57
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v165
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v167
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v176
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v43
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v44
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v45
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v164
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v163
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v177
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v178
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v183
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v42
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v41
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v40
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v148
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v179
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v149
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v179
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v182
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v147
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v146
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v150
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v151
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v167
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v178
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v176
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v177
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v118
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v160
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v161
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v163
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v164
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v165
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v103
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v132
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v133
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v151
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v160
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v161
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v102
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v101
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v134
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v135
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v147
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v150
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v149
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v148
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v100
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v86
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v144
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v119
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v135
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v146
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v144
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v85
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v84
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v128
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v129
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v131
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v134
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v132
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v133
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v83
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v130
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v131
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v119
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v130
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v128
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v129
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v81
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v80
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v113
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v114
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v114
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v118
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v116
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v117
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v115
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v116
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v102
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v115
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v113
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v68
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v117
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v87
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v98
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v103
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v100
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v67
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v96
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v97
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v87
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v99
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v96
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v97
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v64
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v98
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v99
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v86
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v85
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v89
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
@@ -97590,8 +92741,18 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s10, 16
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v51
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v93
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v92
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB59_3
@@ -97609,445 +92770,401 @@ define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s4
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
 ; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v86
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v55
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v54
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v52
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v85
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v51
 ; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v84
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v81
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v82
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v38
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v89, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v90, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v91, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v92, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v93, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v69
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v88, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v74, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v75, v12
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v68
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v65
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v64
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v8, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v54
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v49
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v48
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v39
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v37
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v8, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v46
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v181
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v76, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v77, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v78, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v79, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v63, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v72, v8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v73, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v49
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v35
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v9, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v33
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v3, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v93
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v61, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v43, v16
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v44, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v14, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v32
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v92
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v62
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v56
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v47
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v45
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v183
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v182
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v91
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v90
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v89
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v12, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v13, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v88
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v79
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v78
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v162
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v145
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v118
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v57, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v58, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v59, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v60, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v40, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v41, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v42, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v12, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v13, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v77
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v76
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v75
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v74
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v73
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v14, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v72
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v63
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v62
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v61
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v14, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v15, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v59
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v17, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v15
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v58
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v3, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v57
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v56
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v179, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v160, v21
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v161, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v18
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v19, v15
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v166
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v165
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v164
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v163
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v148
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v147
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v146
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v47
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v46
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v45
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v17, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v18, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v44
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v43
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v42
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v100
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v83
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v167, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v176, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v177, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v178, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v149, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v150, v18
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v151, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v17, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v18, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v41
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v40
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v183
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v182
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v181
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v19, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v180
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v178
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v177
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v19, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v20, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v176
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v167
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v22, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v166
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v3, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v165
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v164
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v144, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v130, v26
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v131, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v19, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v24, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v112
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v103
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v102
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v86
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v85
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v163
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v162
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v161
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v22, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v23, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v160
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v151
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v150
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v26, v27
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v69
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v132, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v133, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v134, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v135, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v119, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v128, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v129, v25
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v22, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v23, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v149
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v148
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v147
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v146
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v145
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v24, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v144
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v135
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v134
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v133
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v24, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v25, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v132
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v131
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v27, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v130
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v3, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v128
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v129
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v119
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v113
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v27, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v115
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v117, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v24, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v29, v25
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v81
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v80
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v71
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v70
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v27
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v68
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v67
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v66
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v65
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 3, v64
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v118
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v117
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v114
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v28, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v116
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v30, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v101
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 3, v99
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v28, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v29, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v112
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v103
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v97
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v102
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xff, v30
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v98
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v87
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v28, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v33, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v35, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v36, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v37, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v113, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v114, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v115, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v116, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v87, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v96, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v97, v30
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v98, v31
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v99, v32
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v27
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v2, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v1, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v30, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v34, v35
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v29, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v34, v30
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v31, v32
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-TRUE16-NEXT:  .LBB59_3: ; %end
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:320
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    s_clause 0x1f
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:320
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:352
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:444
+; GFX11-TRUE16-NEXT:    s_clause 0x7
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:448
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:452
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:456
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:460
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:464
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:468
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:472
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:476
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ; GFX11-TRUE16-NEXT:  .LBB59_4:
@@ -105550,1735 +100667,870 @@ define inreg <16 x i64> @bitcast_v64bf16_to_v16i64_scalar(<64 x bfloat> inreg %a
 ; GFX9-NEXT:  .LBB63_4:
 ; GFX9-NEXT:    s_branch .LBB63_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v16i64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:192
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:164
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v120, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v121, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v122, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v123, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v124, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v125, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v126, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v127, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v136, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v137, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v138, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v139, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v140, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v141, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v142, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v143, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v152, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v153, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v154, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v155, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v156, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v157, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v158, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v159, s32 offset:36
-; GFX11-TRUE16-NEXT:    s_clause 0x8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v168, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v169, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v170, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v171, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v172, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v173, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v174, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v175, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v184, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v183, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v182, v9 :: v_dual_mov_b32 v169, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v171, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v176, v3 :: v_dual_mov_b32 v175, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v173, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v181, s28 :: v_dual_mov_b32 v172, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB63_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB63_3
-; GFX11-TRUE16-NEXT:  .LBB63_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s27, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s27, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s26, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s25, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s25, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff0000, v181
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v8, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v8
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v10, v6
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s24, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s23, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v7, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s22, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v11, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s22, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v9, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s21, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v14, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v11
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v16, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s20, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v12, v16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s20, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s19, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v18, v12
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v16, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s18, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v21, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v18
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s17, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v17, v20
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v19, v22
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s17, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v19, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s16, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, v24, v19
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v25, 16, 1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s16, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, v22, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v19
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, 0x400000, v25
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s3, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, v27, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v23
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v26, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v28, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s2, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, v25, v28
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v30, v25, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s1, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, v30, v25
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v31, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, v28, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v31
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v30, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v33, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, 0x400000, v29
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v30
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v29, v32, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v31, v34, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, v29, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, v31, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v31, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v37, v33
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v35, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v33, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v36, v37
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v183
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v35, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v183
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v37, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v182
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v182
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v35, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v36, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v183, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v38, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v39, v36
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v182, v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v177
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v35, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v177
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v32, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v49, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v49, v35
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v37, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v31, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v176
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v31, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v177, v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v50, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v176
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v49, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v50, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v37, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v48, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, v39, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v176, v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v34, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v37, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v39, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, v38, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v181
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v36, v49, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v36, v38
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, 0x400000, v48
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v50, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v39, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, 0x400000, v39
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v51, v48, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, v38, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v37
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v52, v50, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, v51, v48
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v52, v52, v50
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v52, 0x400000, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v181, v39, 16, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
-; GFX11-TRUE16-NEXT:  .LBB63_3: ; %end
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v181 :: v_dual_mov_b32 v19, v175
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v21, v176
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v184, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v175, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v174, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v173, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v172, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v171, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v170, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v169, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v168, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v159, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v158, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v157, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v156, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v155, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v154, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v153, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v152, off, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v143, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v142, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v141, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v140, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v139, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v138, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v137, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v136, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v127, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v126, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v125, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v124, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v123, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v122, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v121, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v120, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:192
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    s_clause 0x8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v177 :: v_dual_mov_b32 v27, v182
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v183 :: v_dual_mov_b32 v31, v178
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v30, v179
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB63_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
-; GFX11-TRUE16-NEXT:    s_branch .LBB63_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v16i64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:288
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:284
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:280
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:276
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:272
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:268
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:264
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:260
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:256
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:252
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:248
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:244
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:240
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:236
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:232
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:228
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:224
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:220
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:216
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:212
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:208
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:204
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:200
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:196
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:192
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:188
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:184
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:180
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:176
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:172
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:168
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:164
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:160
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:156
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:152
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:148
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:144
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:140
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:136
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:132
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:128
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:124
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:120
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:116
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:112
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:108
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:104
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:100
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v136, s32 offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v137, s32 offset:92
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v138, s32 offset:88
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v139, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v140, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v141, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v142, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v143, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v152, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v153, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v154, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v155, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v156, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v157, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v158, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v159, s32 offset:36
-; GFX11-FAKE16-NEXT:    s_clause 0x8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v168, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v169, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v170, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v171, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v172, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v173, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v174, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v175, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v184, s32
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v181, v9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v182, v10 :: v_dual_mov_b32 v169, v7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v176, v6 :: v_dual_mov_b32 v171, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v173, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v175, v1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v183, s28 :: v_dual_mov_b32 v172, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB63_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB63_3
-; GFX11-FAKE16-NEXT:  .LBB63_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s27, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s27, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s26, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s25, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s25, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s24, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff0000, v183
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v8, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v3, v8
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v10, v6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s24, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s23, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v7, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s23, 16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s22, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v11, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s22, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v9, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v10, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s21, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v14, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s21, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v11
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v16, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s20, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, v12, v16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s20, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s19, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v18, v12
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s19, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v16, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s18, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s18, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v18
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s17, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v17, v20
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v19, v22
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s17, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v19, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s16, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v24, v19
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v25, 16, 1
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v22, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v19
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v25
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s3, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, v27, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v23
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v26, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v25, v28, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s2, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, v25, v28
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, 0x400000, v28
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v30, v25, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s1, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, v30, v25
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v31, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, v28, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v31
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v30, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v33, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, 0x400000, v29
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, 0x400000, v30
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v29, v32, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v31, v34, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, v29, v32
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, v31, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v31, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v37, v33
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v35, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v33, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v36, v37
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v182
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v35, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v182
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v37, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v181
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v181
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v35, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v36, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v182, v31, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v38, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v39, v36
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v181, v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v176
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v35, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v176
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v32, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v49, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v37, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v31, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v177
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v31, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v176, v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v177
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v50, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v37, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v48, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, v39, v38
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v177, v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v34, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v37, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v39, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, v38, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v36, v49 :: v_dual_lshlrev_b32 v48, 16, v183
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v36, v38
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v48
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v39, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v39
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v48, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, v38, v39
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v37
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v50, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v51, v51, v48
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v52, v52, v50
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, 0x400000, v50
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v183, v39, 16, v48
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
-; GFX11-FAKE16-NEXT:  .LBB63_3: ; %end
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v19, v175
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v184, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v175, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v174, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v173, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v172, off, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v171, off, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v170, off, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v169, off, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v168, off, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v159, off, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v158, off, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v157, off, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v156, off, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v155, off, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v154, off, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v153, off, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v152, off, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v143, off, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v142, off, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v141, off, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v140, off, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v139, off, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v138, off, s32 offset:88
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v137, off, s32 offset:92
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v136, off, s32 offset:96
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:100
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:104
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:108
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:112
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:116
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:120
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:124
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:128
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:132
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:136
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:140
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:144
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:148
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:152
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:156
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:160
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:164
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:168
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:172
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:176
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:180
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:184
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:188
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:192
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:196
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:200
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:204
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:208
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:212
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:216
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:220
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:224
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:228
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:232
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:236
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:240
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:244
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:248
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:252
-; GFX11-FAKE16-NEXT:    s_clause 0x8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:256
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:260
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:264
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:268
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:272
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:276
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:280
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:284
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:288
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v183 :: v_dual_mov_b32 v21, v177
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, v176 :: v_dual_mov_b32 v27, v181
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v28, v182
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, v179 :: v_dual_mov_b32 v31, v178
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB63_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
-; GFX11-FAKE16-NEXT:    s_branch .LBB63_2
+; GFX11-LABEL: bitcast_v64bf16_to_v16i64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:288
+; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:284
+; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:280
+; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:276
+; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:272
+; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:268
+; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:264
+; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:260
+; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:256
+; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:252
+; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:248
+; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:244
+; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:240
+; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:236
+; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:232
+; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:228
+; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:224
+; GFX11-NEXT:    scratch_store_b32 off, v73, s32 offset:220
+; GFX11-NEXT:    scratch_store_b32 off, v74, s32 offset:216
+; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:212
+; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:208
+; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:204
+; GFX11-NEXT:    scratch_store_b32 off, v78, s32 offset:200
+; GFX11-NEXT:    scratch_store_b32 off, v79, s32 offset:196
+; GFX11-NEXT:    scratch_store_b32 off, v88, s32 offset:192
+; GFX11-NEXT:    scratch_store_b32 off, v89, s32 offset:188
+; GFX11-NEXT:    scratch_store_b32 off, v90, s32 offset:184
+; GFX11-NEXT:    scratch_store_b32 off, v91, s32 offset:180
+; GFX11-NEXT:    scratch_store_b32 off, v92, s32 offset:176
+; GFX11-NEXT:    scratch_store_b32 off, v93, s32 offset:172
+; GFX11-NEXT:    scratch_store_b32 off, v94, s32 offset:168
+; GFX11-NEXT:    scratch_store_b32 off, v95, s32 offset:164
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_store_b32 off, v104, s32 offset:160
+; GFX11-NEXT:    scratch_store_b32 off, v105, s32 offset:156
+; GFX11-NEXT:    scratch_store_b32 off, v106, s32 offset:152
+; GFX11-NEXT:    scratch_store_b32 off, v107, s32 offset:148
+; GFX11-NEXT:    scratch_store_b32 off, v108, s32 offset:144
+; GFX11-NEXT:    scratch_store_b32 off, v109, s32 offset:140
+; GFX11-NEXT:    scratch_store_b32 off, v110, s32 offset:136
+; GFX11-NEXT:    scratch_store_b32 off, v111, s32 offset:132
+; GFX11-NEXT:    scratch_store_b32 off, v120, s32 offset:128
+; GFX11-NEXT:    scratch_store_b32 off, v121, s32 offset:124
+; GFX11-NEXT:    scratch_store_b32 off, v122, s32 offset:120
+; GFX11-NEXT:    scratch_store_b32 off, v123, s32 offset:116
+; GFX11-NEXT:    scratch_store_b32 off, v124, s32 offset:112
+; GFX11-NEXT:    scratch_store_b32 off, v125, s32 offset:108
+; GFX11-NEXT:    scratch_store_b32 off, v126, s32 offset:104
+; GFX11-NEXT:    scratch_store_b32 off, v127, s32 offset:100
+; GFX11-NEXT:    scratch_store_b32 off, v136, s32 offset:96
+; GFX11-NEXT:    scratch_store_b32 off, v137, s32 offset:92
+; GFX11-NEXT:    scratch_store_b32 off, v138, s32 offset:88
+; GFX11-NEXT:    scratch_store_b32 off, v139, s32 offset:84
+; GFX11-NEXT:    scratch_store_b32 off, v140, s32 offset:80
+; GFX11-NEXT:    scratch_store_b32 off, v141, s32 offset:76
+; GFX11-NEXT:    scratch_store_b32 off, v142, s32 offset:72
+; GFX11-NEXT:    scratch_store_b32 off, v143, s32 offset:68
+; GFX11-NEXT:    scratch_store_b32 off, v152, s32 offset:64
+; GFX11-NEXT:    scratch_store_b32 off, v153, s32 offset:60
+; GFX11-NEXT:    scratch_store_b32 off, v154, s32 offset:56
+; GFX11-NEXT:    scratch_store_b32 off, v155, s32 offset:52
+; GFX11-NEXT:    scratch_store_b32 off, v156, s32 offset:48
+; GFX11-NEXT:    scratch_store_b32 off, v157, s32 offset:44
+; GFX11-NEXT:    scratch_store_b32 off, v158, s32 offset:40
+; GFX11-NEXT:    scratch_store_b32 off, v159, s32 offset:36
+; GFX11-NEXT:    s_clause 0x8
+; GFX11-NEXT:    scratch_store_b32 off, v168, s32 offset:32
+; GFX11-NEXT:    scratch_store_b32 off, v169, s32 offset:28
+; GFX11-NEXT:    scratch_store_b32 off, v170, s32 offset:24
+; GFX11-NEXT:    scratch_store_b32 off, v171, s32 offset:20
+; GFX11-NEXT:    scratch_store_b32 off, v172, s32 offset:16
+; GFX11-NEXT:    scratch_store_b32 off, v173, s32 offset:12
+; GFX11-NEXT:    scratch_store_b32 off, v174, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v175, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v184, s32
+; GFX11-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
+; GFX11-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v181, v9
+; GFX11-NEXT:    v_dual_mov_b32 v182, v10 :: v_dual_mov_b32 v169, v7
+; GFX11-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v3
+; GFX11-NEXT:    v_dual_mov_b32 v176, v6 :: v_dual_mov_b32 v171, v4
+; GFX11-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v173, v0
+; GFX11-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v175, v1
+; GFX11-NEXT:    v_dual_mov_b32 v183, s28 :: v_dual_mov_b32 v172, s29
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB63_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
+; GFX11-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
+; GFX11-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
+; GFX11-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
+; GFX11-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
+; GFX11-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
+; GFX11-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
+; GFX11-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB63_3
+; GFX11-NEXT:  .LBB63_2: ; %cmp.true
+; GFX11-NEXT:    s_and_b32 s5, s27, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s4, s27, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s6, s26, 16
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v1
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_bfe_u32 v10, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_bfe_u32 v9, v3, 16, 1
+; GFX11-NEXT:    s_lshl_b32 s7, s25, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    s_and_b32 s5, s25, 0xffff0000
+; GFX11-NEXT:    s_and_b32 s4, s24, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
+; GFX11-NEXT:    v_and_b32_e32 v51, 0xffff0000, v183
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v10, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v9, v3
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_bfe_u32 v10, v6, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_bfe_u32 v3, v8, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v3, v8
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v10, v6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    s_lshl_b32 s4, s24, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v6
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v8
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
+; GFX11-NEXT:    s_and_b32 s4, s23, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v4, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v5
+; GFX11-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
+; GFX11-NEXT:    v_bfe_u32 v7, v9, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, v7, v9
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s23, 16
+; GFX11-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_bfe_u32 v11, v7, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    s_and_b32 s4, s22, 0xffff0000
+; GFX11-NEXT:    v_bfe_u32 v9, v12, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v11, v7
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s22, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
+; GFX11-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v9, v12
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v7
+; GFX11-NEXT:    v_bfe_u32 v14, v10, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_or_b32_e32 v13, 0x400000, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
+; GFX11-NEXT:    s_and_b32 s4, s21, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v9, v11, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, v14, v10
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
+; GFX11-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v10
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    s_lshl_b32 s4, s21, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v11
+; GFX11-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v10, v13, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_bfe_u32 v12, v16, 16, 1
+; GFX11-NEXT:    s_and_b32 s4, s20, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, v12, v16
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    s_lshl_b32 s4, s20, 16
+; GFX11-NEXT:    v_or_b32_e32 v17, 0x400000, v16
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
+; GFX11-NEXT:    v_bfe_u32 v18, v12, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
+; GFX11-NEXT:    s_and_b32 s4, s19, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v18, v12
+; GFX11-NEXT:    v_bfe_u32 v16, v19, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v9
+; GFX11-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s19, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v16, v19
+; GFX11-NEXT:    v_or_b32_e32 v16, 0x400000, v12
+; GFX11-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v21, v17, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
+; GFX11-NEXT:    v_or_b32_e32 v20, 0x400000, v19
+; GFX11-NEXT:    s_and_b32 s4, s18, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v16, v18, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
+; GFX11-NEXT:    v_add_nc_u32_e32 v19, v21, v17
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
+; GFX11-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
+; GFX11-NEXT:    v_or_b32_e32 v19, 0x400000, v17
+; GFX11-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
+; GFX11-NEXT:    s_lshl_b32 s4, s18, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
+; GFX11-NEXT:    v_or_b32_e32 v21, 0x400000, v18
+; GFX11-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v17, v20, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_bfe_u32 v19, v22, 16, 1
+; GFX11-NEXT:    s_and_b32 s4, s17, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v17, v17, v20
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
+; GFX11-NEXT:    v_or_b32_e32 v21, 0x400000, v20
+; GFX11-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
+; GFX11-NEXT:    v_add_nc_u32_e32 v18, v19, v22
+; GFX11-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
+; GFX11-NEXT:    s_lshl_b32 s4, s17, 16
+; GFX11-NEXT:    v_or_b32_e32 v23, 0x400000, v22
+; GFX11-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
+; GFX11-NEXT:    v_bfe_u32 v24, v19, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
+; GFX11-NEXT:    s_and_b32 s4, s16, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, v24, v19
+; GFX11-NEXT:    v_bfe_u32 v22, v25, 16, 1
+; GFX11-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
+; GFX11-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s16, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
+; GFX11-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, v22, v25
+; GFX11-NEXT:    v_or_b32_e32 v22, 0x400000, v19
+; GFX11-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v27, v23, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
+; GFX11-NEXT:    v_or_b32_e32 v26, 0x400000, v25
+; GFX11-NEXT:    s_and_b32 s4, s3, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v22, v24, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
+; GFX11-NEXT:    v_add_nc_u32_e32 v25, v27, v23
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
+; GFX11-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
+; GFX11-NEXT:    v_or_b32_e32 v25, 0x400000, v23
+; GFX11-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
+; GFX11-NEXT:    v_or_b32_e32 v27, 0x400000, v24
+; GFX11-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
+; GFX11-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v23, v26, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
+; GFX11-NEXT:    v_bfe_u32 v25, v28, 16, 1
+; GFX11-NEXT:    s_and_b32 s3, s2, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
+; GFX11-NEXT:    v_or_b32_e32 v27, 0x400000, v26
+; GFX11-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v24, v25, v28
+; GFX11-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_or_b32_e32 v29, 0x400000, v28
+; GFX11-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
+; GFX11-NEXT:    v_bfe_u32 v30, v25, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
+; GFX11-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
+; GFX11-NEXT:    s_and_b32 s2, s1, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, v30, v25
+; GFX11-NEXT:    v_bfe_u32 v28, v31, 16, 1
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, v28, v31
+; GFX11-NEXT:    v_or_b32_e32 v28, 0x400000, v25
+; GFX11-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
+; GFX11-NEXT:    v_bfe_u32 v33, v29, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
+; GFX11-NEXT:    v_or_b32_e32 v32, 0x400000, v31
+; GFX11-NEXT:    s_and_b32 s1, s0, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v28, v30, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v33, v29
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
+; GFX11-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
+; GFX11-NEXT:    v_or_b32_e32 v31, 0x400000, v29
+; GFX11-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
+; GFX11-NEXT:    v_or_b32_e32 v33, 0x400000, v30
+; GFX11-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
+; GFX11-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v29, v32, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
+; GFX11-NEXT:    v_or_b32_e32 v35, 0x400000, v32
+; GFX11-NEXT:    v_bfe_u32 v31, v34, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
+; GFX11-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, v29, v32
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
+; GFX11-NEXT:    v_add_nc_u32_e32 v30, v31, v34
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
+; GFX11-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
+; GFX11-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
+; GFX11-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v37, v33, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v29
+; GFX11-NEXT:    v_bfe_u32 v35, v31, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v37, v33
+; GFX11-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v35, v31
+; GFX11-NEXT:    v_or_b32_e32 v37, 0x400000, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_or_b32_e32 v38, 0x400000, v31
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v34, v36, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v33, v35, 16, 1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v33, v35
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_bfe_u32 v36, v37, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_bfe_u32 v35, v38, 16, 1
+; GFX11-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v36, v37
+; GFX11-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v182
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v35, v38
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v182
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xffff, v30
+; GFX11-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_bfe_u32 v37, v36, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v37, v36
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v181
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v181
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-NEXT:    v_bfe_u32 v35, v37, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v36, v38, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v35, v37
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v36, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v37
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v39, v36, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_lshl_or_b32 v182, v31, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v38, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v39, v36
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v38, 0x400000, v36
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
+; GFX11-NEXT:    v_lshl_or_b32 v181, v32, 16, v33
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff0000, v176
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_bfe_u32 v35, v37, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
+; GFX11-NEXT:    v_bfe_u32 v32, v36, 16, 1
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v35, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 16, v176
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v32, v36
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_bfe_u32 v37, v38, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v49, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v49, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v37, v35, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
+; GFX11-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v37, v35
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v31, v36, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v177
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v31, v36
+; GFX11-NEXT:    v_lshl_or_b32 v176, v33, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
+; GFX11-NEXT:    v_or_b32_e32 v34, 0x400000, v35
+; GFX11-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v34, 0x400000, v36
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
+; GFX11-NEXT:    v_bfe_u32 v37, v32, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v32
+; GFX11-NEXT:    v_bfe_u32 v50, v38, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
+; GFX11-NEXT:    v_bfe_u32 v34, v35, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v177
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_bfe_u32 v49, v37, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v50, v38
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v37
+; GFX11-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_bfe_u32 v48, v37, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v39, v38, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
+; GFX11-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v48, v37
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
+; GFX11-NEXT:    v_add_nc_u32_e32 v39, v39, v38
+; GFX11-NEXT:    v_lshl_or_b32 v177, v35, 16, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_or_b32_e32 v35, 0x400000, v37
+; GFX11-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v37, v33, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v39, v34, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v37, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v39, v34
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, v38, v35
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v33, v36, v49 :: v_dual_lshlrev_b32 v48, 16, v183
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v36, v38, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v38
+; GFX11-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v36, v38
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_or_b32_e32 v55, 0x400000, v48
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_bfe_u32 v50, v37, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v38, v39, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v54, 0x400000, v39
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
+; GFX11-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
+; GFX11-NEXT:    v_bfe_u32 v51, v48, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, v38, v39
+; GFX11-NEXT:    v_or_b32_e32 v53, 0x400000, v37
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
+; GFX11-NEXT:    v_bfe_u32 v52, v50, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v51, v51, v48
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v52, v52, v50
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
+; GFX11-NEXT:    v_or_b32_e32 v52, 0x400000, v50
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
+; GFX11-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
+; GFX11-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
+; GFX11-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
+; GFX11-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
+; GFX11-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
+; GFX11-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
+; GFX11-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
+; GFX11-NEXT:    v_lshl_or_b32 v183, v39, 16, v48
+; GFX11-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
+; GFX11-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
+; GFX11-NEXT:  .LBB63_3: ; %end
+; GFX11-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
+; GFX11-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
+; GFX11-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
+; GFX11-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
+; GFX11-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
+; GFX11-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
+; GFX11-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v19, v175
+; GFX11-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
+; GFX11-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
+; GFX11-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_load_b32 v184, off, s32
+; GFX11-NEXT:    scratch_load_b32 v175, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v174, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v173, off, s32 offset:12
+; GFX11-NEXT:    scratch_load_b32 v172, off, s32 offset:16
+; GFX11-NEXT:    scratch_load_b32 v171, off, s32 offset:20
+; GFX11-NEXT:    scratch_load_b32 v170, off, s32 offset:24
+; GFX11-NEXT:    scratch_load_b32 v169, off, s32 offset:28
+; GFX11-NEXT:    scratch_load_b32 v168, off, s32 offset:32
+; GFX11-NEXT:    scratch_load_b32 v159, off, s32 offset:36
+; GFX11-NEXT:    scratch_load_b32 v158, off, s32 offset:40
+; GFX11-NEXT:    scratch_load_b32 v157, off, s32 offset:44
+; GFX11-NEXT:    scratch_load_b32 v156, off, s32 offset:48
+; GFX11-NEXT:    scratch_load_b32 v155, off, s32 offset:52
+; GFX11-NEXT:    scratch_load_b32 v154, off, s32 offset:56
+; GFX11-NEXT:    scratch_load_b32 v153, off, s32 offset:60
+; GFX11-NEXT:    scratch_load_b32 v152, off, s32 offset:64
+; GFX11-NEXT:    scratch_load_b32 v143, off, s32 offset:68
+; GFX11-NEXT:    scratch_load_b32 v142, off, s32 offset:72
+; GFX11-NEXT:    scratch_load_b32 v141, off, s32 offset:76
+; GFX11-NEXT:    scratch_load_b32 v140, off, s32 offset:80
+; GFX11-NEXT:    scratch_load_b32 v139, off, s32 offset:84
+; GFX11-NEXT:    scratch_load_b32 v138, off, s32 offset:88
+; GFX11-NEXT:    scratch_load_b32 v137, off, s32 offset:92
+; GFX11-NEXT:    scratch_load_b32 v136, off, s32 offset:96
+; GFX11-NEXT:    scratch_load_b32 v127, off, s32 offset:100
+; GFX11-NEXT:    scratch_load_b32 v126, off, s32 offset:104
+; GFX11-NEXT:    scratch_load_b32 v125, off, s32 offset:108
+; GFX11-NEXT:    scratch_load_b32 v124, off, s32 offset:112
+; GFX11-NEXT:    scratch_load_b32 v123, off, s32 offset:116
+; GFX11-NEXT:    scratch_load_b32 v122, off, s32 offset:120
+; GFX11-NEXT:    scratch_load_b32 v121, off, s32 offset:124
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_load_b32 v120, off, s32 offset:128
+; GFX11-NEXT:    scratch_load_b32 v111, off, s32 offset:132
+; GFX11-NEXT:    scratch_load_b32 v110, off, s32 offset:136
+; GFX11-NEXT:    scratch_load_b32 v109, off, s32 offset:140
+; GFX11-NEXT:    scratch_load_b32 v108, off, s32 offset:144
+; GFX11-NEXT:    scratch_load_b32 v107, off, s32 offset:148
+; GFX11-NEXT:    scratch_load_b32 v106, off, s32 offset:152
+; GFX11-NEXT:    scratch_load_b32 v105, off, s32 offset:156
+; GFX11-NEXT:    scratch_load_b32 v104, off, s32 offset:160
+; GFX11-NEXT:    scratch_load_b32 v95, off, s32 offset:164
+; GFX11-NEXT:    scratch_load_b32 v94, off, s32 offset:168
+; GFX11-NEXT:    scratch_load_b32 v93, off, s32 offset:172
+; GFX11-NEXT:    scratch_load_b32 v92, off, s32 offset:176
+; GFX11-NEXT:    scratch_load_b32 v91, off, s32 offset:180
+; GFX11-NEXT:    scratch_load_b32 v90, off, s32 offset:184
+; GFX11-NEXT:    scratch_load_b32 v89, off, s32 offset:188
+; GFX11-NEXT:    scratch_load_b32 v88, off, s32 offset:192
+; GFX11-NEXT:    scratch_load_b32 v79, off, s32 offset:196
+; GFX11-NEXT:    scratch_load_b32 v78, off, s32 offset:200
+; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:204
+; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:208
+; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:212
+; GFX11-NEXT:    scratch_load_b32 v74, off, s32 offset:216
+; GFX11-NEXT:    scratch_load_b32 v73, off, s32 offset:220
+; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:224
+; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:228
+; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:232
+; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:236
+; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:240
+; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:244
+; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:248
+; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:252
+; GFX11-NEXT:    s_clause 0x8
+; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:256
+; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:260
+; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:264
+; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:268
+; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:272
+; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:276
+; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:280
+; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:284
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:288
+; GFX11-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
+; GFX11-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
+; GFX11-NEXT:    v_dual_mov_b32 v16, v183 :: v_dual_mov_b32 v21, v177
+; GFX11-NEXT:    v_dual_mov_b32 v24, v176 :: v_dual_mov_b32 v27, v181
+; GFX11-NEXT:    v_mov_b32_e32 v28, v182
+; GFX11-NEXT:    v_dual_mov_b32 v30, v179 :: v_dual_mov_b32 v31, v178
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB63_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
+; GFX11-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
+; GFX11-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
+; GFX11-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
+; GFX11-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
+; GFX11-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
+; GFX11-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
+; GFX11-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
+; GFX11-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
+; GFX11-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
+; GFX11-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
+; GFX11-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
+; GFX11-NEXT:    s_branch .LBB63_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -122482,2068 +116734,1024 @@ define inreg <128 x i8> @bitcast_v16f64_to_v128i8_scalar(<16 x double> inreg %a,
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v16f64_to_v128i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:84
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s30, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s96, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s97, 1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s98, 2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s99, 3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s100, 4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s101, 5
-; GFX11-TRUE16-NEXT:    s_mov_b32 vcc_hi, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_clause 0x11
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s102, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s103, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s104, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s50, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s51, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s52, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s53, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s54, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s55, 15
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s64, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s65, 17
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s66, 18
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s67, 19
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s68, 20
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s69, 21
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s70, 22
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s71, 23
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s80, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s81, 25
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s82, 26
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s83, 27
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s84, 28
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s85, 29
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s86, 30
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s87, 31
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB73_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s51, s40, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s52, s40, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s53, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s54, s14, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s41, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s55, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s64, s12, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s65, s10, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s66, s10, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s67, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s68, s8, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s69, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s70, s6, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s71, s4, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s15, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s80, s4, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s81, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s82, s28, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s83, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s84, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s85, s24, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s86, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s87, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s96, s22, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s13, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s97, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s98, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s18, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s100, s18, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 31
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s101, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s16, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 30
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s11, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s3, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 29
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s1, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 28
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s0, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 27
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s9, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 26
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[12:13], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[76:77], s[10:11], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 25
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[88:89], s[6:7], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[90:91], s[4:5], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[28:29], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s7, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 23
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 22
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 21
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 20
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s29, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 19
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s29, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 18
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s29, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 17
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 15
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 14
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 13
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 11
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB73_4
-; GFX11-TRUE16-NEXT:  .LBB73_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[13:14], s[8:9], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[1:2], s[22:23], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[50:51], s[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[15:16], s[6:7], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[3:4], s[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[54:55], s[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[5:6], s[40:41], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[7:8], s[14:15], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[9:10], s[12:13], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[11:12], s[10:11], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], s[4:5], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[23:24], s[28:29], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[27:28], s[26:27], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[31:32], s[24:25], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[37:38], s[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[66:67], s[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[80:81], 24, v[50:51]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[15:16]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[69:70], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[81:82], 24, v[54:55]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[5:6]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[25:26], 24, v[9:10]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[29:30], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[19:20]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[23:24]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[27:28]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[31:32]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[37:38]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[82:83], 24, v[66:67]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 24, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 8, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v163, 24, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v176, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v181, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v40, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v166, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v167, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v42, 24, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v44, 16, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v46, 8, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v179, 16, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v180, 8, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v57, 24, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v58, 16, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v183, 16, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v41, 8, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v60, 24, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v54
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v45, 8, v54
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v67
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v67
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v67
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v66
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v66
-; GFX11-TRUE16-NEXT:    s_branch .LBB73_5
-; GFX11-TRUE16-NEXT:  .LBB73_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr104_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr103_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr102_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr101_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr99_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr98_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr97_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr96_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr87_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr86_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr85_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr84_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 2
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 8
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 9
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 10
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 11
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 12
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 13
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 14
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 15
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 17
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 18
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 19
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 20
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 21
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 22
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 23
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 24
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 25
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 26
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 27
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 28
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 29
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 30
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s43, 31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 2
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s43, 8
-; GFX11-TRUE16-NEXT:    s_branch .LBB73_2
-; GFX11-TRUE16-NEXT:  .LBB73_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v71.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, s34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v56.l, s35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v43.l, s103
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, s104
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v183.l, s101
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v96.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 31
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v41.l, s102
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v179.l, s99
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v180.l, s100
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v166.l, s97
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v99.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 30
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v167.l, s98
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v161.l, s87
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v162.l, s96
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v148.l, s85
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v100.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v149.l, s86
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v135.l, s83
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v147.l, s84
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v132.l, s81
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v101.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v133.l, s82
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v119.l, s71
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v129.l, s80
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v115.l, s69
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v112.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v116.l, s70
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v102.l, s67
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v103.l, s68
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v97.l, s65
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v113.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v98.l, s66
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.l, s55
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v86.l, s64
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s53
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v114.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, s54
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s52
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v59.l, s36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v117.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, s37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v61.l, s38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, s39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v63.l, s48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v118.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v72.l, s49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v73.l, s50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v82.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v81.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v80.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v70.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v69.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v68.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v130.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, s30
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s94
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s92
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s90
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v131.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s88
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s78
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s76
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s74
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v134.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s72
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s62
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v144.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v145.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 17
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v146.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v150.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 15
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v151.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v160.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 13
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v163.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v164.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 11
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v165.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v176.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v177.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v178.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v181.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v182.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v44.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v46.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, s0
-; GFX11-TRUE16-NEXT:  .LBB73_5: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v56
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xff, v66
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v66, v83
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v83, 0xff, v47
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v47, 8, v63
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v45, 8, v45
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v43, 0xff, v43
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v82, v83, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v73
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v45, v54, v45
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, v43, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v43, 0xff, v61
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 16, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, v67, v83
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v83, 0xff, v72
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xff, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v183, 0xff, v183
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, v66, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, v83, v47
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xff, v55
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v47, 8, v60
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v80
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v62
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v80, v183, v80
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, v66, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v45
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, v82, v83
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v82, v43, v47
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v41
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v56, v66, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v81
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v50, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v59
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xff, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v57
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v57, v66, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v51, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, v82, v83
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xff, v179
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v80, v50, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v82, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v46
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xff, v44
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v42
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, v37, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, v50, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v38, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v70, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v167
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v166
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v40
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v67, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v82, v37, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, v38, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xff, v182
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v66
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v181
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v162
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xff, v161
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v178
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v177
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v176
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v38, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v66, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v69, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[54:57], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[80:83], off offset:16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v66, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v67, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v149
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xff, v148
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v165
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xff, v164
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 8, v163
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v147
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v31, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, v38, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v32, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v54, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v27, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xff, v135
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v160
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xff, v151
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 8, v150
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v133
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xff, v132
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v50, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v28, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v54, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v64
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, v66, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v31, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, v32, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v68, v27, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, v28, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v80, v23, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v146
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v145
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v144
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v129
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v119
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v134
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v27, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v19, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v32, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v131
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v130
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v116
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xff, v115
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v128
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xff, v118
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v117
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v28, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v35, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v38, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, v23, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v82, v19, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, v20, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v15, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v103
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v102
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v114
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v113
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v112
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v98
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v16, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v23, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v97
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v101
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v100
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v99
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v86
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v19, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v24, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v29, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, v13, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, v14, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v9, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v87
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v85
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v49
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v71
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v14, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v19, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v53
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v20, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v22, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v19, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v20, v17
-; GFX11-TRUE16-NEXT:    s_clause 0x5
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[66:69], off offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[80:83], off offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[31:34], off offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[11:14], off offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:112
-; GFX11-TRUE16-NEXT:    s_clause 0x11
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s104, v75, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s103, v75, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s102, v75, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s101, v75, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s100, v75, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s99, v75, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s98, v75, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s97, v75, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s96, v75, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s87, v74, 31
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s86, v74, 30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s85, v74, 29
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s84, v74, 28
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s83, v74, 27
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s82, v74, 26
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s81, v74, 25
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s80, v74, 24
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s71, v74, 23
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s70, v74, 22
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s69, v74, 21
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s68, v74, 20
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v74, 19
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v74, 18
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v74, 17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v74, 16
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v74, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v74, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v74, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v74, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v74, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v74, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v74, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v74, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v74, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v74, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v74, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v74, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v74, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v74, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v74, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v74, 0
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v16f64_to_v128i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:88
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:92
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s96, 0
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s97, 1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s98, 2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s99, 3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s100, 4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s101, 5
-; GFX11-FAKE16-NEXT:    s_mov_b32 vcc_hi, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_clause 0x13
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s102, 6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr79 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s103, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s104, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s49, 9
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s50, 10
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s51, 11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s52, 12
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s53, 13
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s54, 14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s55, 15
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s64, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s65, 17
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s66, 18
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s67, 19
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s68, 20
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s69, 21
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s70, 22
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s71, 23
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s80, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s81, 25
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s82, 26
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s83, 27
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s84, 28
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s85, 29
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s86, 30
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s87, 31
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB73_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s50, s41, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s41, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s41, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s52, s40, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s11, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s51, s40, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s15, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s15, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s15, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s54, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s53, s14, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s11, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s13, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s64, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s55, s12, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s66, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s65, s10, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s68, s8, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s67, s8, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s70, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s69, s6, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s80, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s71, s4, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s82, s28, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s7, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s81, s28, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s84, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s83, s26, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s86, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s85, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s96, s22, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s7, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s87, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s98, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s97, s20, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 10
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s99, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s16, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 11
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s101, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s2, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 13
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s29, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[74:75], s[12:13], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[76:77], s[10:11], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 14
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[88:89], s[6:7], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[90:91], s[4:5], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 15
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s29, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[92:93], s[28:29], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 17
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 18
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 19
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 20
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 21
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 22
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 23
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 25
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 26
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 27
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 28
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 29
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 30
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 31
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v79, s42, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v79, s42, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v79, s42, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v79, s42, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v79, s42, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v79, s42, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v79, s42, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v79, s42, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v79, s42, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB73_4
-; GFX11-FAKE16-NEXT:  .LBB73_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[23:24], s[24:25], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[28:29], s[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[32:33], s[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[5:6], s[12:13], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[36:37], s[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[52:53], s[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[1:2], s[40:41], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[3:4], s[14:15], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[7:8], s[10:11], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[9:10], s[8:9], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[11:12], s[6:7], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[13:14], s[4:5], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[15:16], s[28:29], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[19:20], s[26:27], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[48:49], s[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[64:65], s[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[23:24]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[28:29]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[32:33]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[25:26], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[69:70], 24, v[36:37]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[80:81], 24, v[52:53]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[21:22], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[7:8]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[30:31], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[11:12]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[19:20]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[48:49]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[81:82], 24, v[64:65]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 24, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 24, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 24, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 8, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 24, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 8, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 8, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 24, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 8, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 24, v49
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 16, v49
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v49
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 24, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v52
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v52
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v65
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v65
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v65
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 16, v64
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 8, v64
-; GFX11-FAKE16-NEXT:    s_branch .LBB73_5
-; GFX11-FAKE16-NEXT:  .LBB73_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr103
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr104
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr101
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr102
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr99
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr100
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr97
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr98
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr87
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr96
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr85
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr86
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
-; GFX11-FAKE16-NEXT:    s_branch .LBB73_2
-; GFX11-FAKE16-NEXT:  .LBB73_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v64, s0 :: v_dual_mov_b32 v65, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s40 :: v_dual_mov_b32 v2, s41
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s14 :: v_dual_mov_b32 v4, s15
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v87, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v39, s54
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s12 :: v_dual_mov_b32 v6, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v96, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s8 :: v_dual_mov_b32 v10, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s6 :: v_dual_mov_b32 v12, s7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v99, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v55, s53
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s4 :: v_dual_mov_b32 v14, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v100, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s26 :: v_dual_mov_b32 v20, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s24 :: v_dual_mov_b32 v24, s25
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v101, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s51 :: v_dual_mov_b32 v28, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v29, s23 :: v_dual_mov_b32 v32, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v33, s21 :: v_dual_mov_b32 v112, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v31, s49 :: v_dual_mov_b32 v36, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v37, s19 :: v_dual_mov_b32 v48, s16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v49, s17 :: v_dual_mov_b32 v114, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v35, s48 :: v_dual_mov_b32 v52, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v53, s3 :: v_dual_mov_b32 v44, s35
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v41, s104 :: v_dual_mov_b32 v116, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v46, s34 :: v_dual_mov_b32 v43, s103
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v181, s102 :: v_dual_mov_b32 v182, s101
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v119, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v51, s39 :: v_dual_mov_b32 v176, s100
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v177, s99
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v163, s98 :: v_dual_mov_b32 v160, s96
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v128, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v165, s97 :: v_dual_mov_b32 v148, s86
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v161, s87 :: v_dual_mov_b32 v144, s83
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v129, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 11
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v71, s38
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v149, s85 :: v_dual_mov_b32 v130, s82
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v135, s84 :: v_dual_mov_b32 v118, s71
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v132, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v131, s81 :: v_dual_mov_b32 v102, s68
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v117, s80 :: v_dual_mov_b32 v98, s65
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v133, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v83, s37
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v113, s70 :: v_dual_mov_b32 v84, s64
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v115, s69 :: v_dual_mov_b32 v86, s55
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v134, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v103, s67 :: v_dual_mov_b32 v18, s52
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v97, s66 :: v_dual_mov_b32 v22, s50
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v145, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 15
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v85, s36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v81, s42 :: v_dual_mov_b32 v38, s90
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v69, s56 :: v_dual_mov_b32 v34, s88
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v146, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v67, s60 :: v_dual_mov_b32 v30, s78
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s76 :: v_dual_mov_b32 v25, s74
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v147, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 17
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v21, s72
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s62 :: v_dual_mov_b32 v80, s44
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v70, s46
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v150, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 18
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v68, s58
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v66, s30
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v54, s94
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, s92 :: v_dual_mov_b32 v151, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 19
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v162, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 20
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v164, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 21
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v166, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 22
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v167, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 23
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v178, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v179, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 25
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v180, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 26
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v183, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v40, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 28
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v42, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 29
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v45, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 30
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v47, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 31
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v56, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v79, 0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v57, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v79, 1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v58, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v79, 2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v59, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v79, 3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v60, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v79, 4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v61, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v79, 5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v62, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v79, 6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v63, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v79, 7
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v72, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v79, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v73, s0
-; GFX11-FAKE16-NEXT:  .LBB73_5: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v46
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xff, v64
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v81
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xff, v65
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v80
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xff, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, v64, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xff, v44
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v44, 8, v63
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v43, 8, v43
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v41, 0xff, v41
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xffff, v64
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v81, v82, v81
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v73
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v52, v43
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v80, v41, v80
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xff, v53
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 16, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, v65, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xff, v72
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v41, 8, v60
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v72, v64, v81
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xffff, v65
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, v82, v44
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v62
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xff, v48
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v49, 0xff, v49
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xff, v61
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v53, v81
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v81, 0xff, v181
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xff, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v73, v64, v65
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v80
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, v82, v41
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v182
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v69
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v74, v52, v64
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v53
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v65
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, v48, v80
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, v81, v70
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v59
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v70, 0xff, v58
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v57
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v75, v52, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v64
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, v49, v65
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v70, v80
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v177
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xff, v176
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v43, v48, v52
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v49
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v53
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v36, v64
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v65, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v56
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xff, v47
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v45
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v44, v48, v49
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v37, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, v64, v65
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v165
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xff, v163
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v68
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v33
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v42
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v32, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v53, v64
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v33, v65
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v45, v36, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v46, v37, v49
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xff, v40
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v52
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v183
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 8, v161
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xff, v160
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v67
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v180
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xff, v179
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v178
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v37, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v49
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, v52, v53
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v29, v64
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, v65, v67
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, v32, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v68, v33, v37
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v28, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v70, v29, v49
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v149
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v148
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v167
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xff, v166
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v164
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v144
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v23, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v36, v37
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v19, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v135
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v54
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v162
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xff, v151
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v150
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 8, v131
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xff, v130
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v50
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v32, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v37, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v49
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v52, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v148, v23, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v150, v19, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v151, v20, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v130, v15, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v147
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v146
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v145
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v118
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v149, v24, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v117
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v134
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v19, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v133
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v132
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v24, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v115
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v113
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v129
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v128
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v119
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v24
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v28, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v33, v34
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v131, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v132, v13, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v133, v14, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v103
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v102
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v30
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v116
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v114
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v112
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v98
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v19, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v97
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v101
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v86
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v84
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v100
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v99
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v24
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v26, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v23
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v9, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v10, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v5, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v96
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v87
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v85
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v39
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v55
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v83
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v10, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v19, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v51
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v71
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v31
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v18, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v21, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v5, v6
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[72:75], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[43:46], off offset:16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v19, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v20, v17
-; GFX11-FAKE16-NEXT:    s_clause 0x5
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[67:70], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[148:151], off offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[130:133], off offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
-; GFX11-FAKE16-NEXT:    s_clause 0x13
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:76
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s104, v77, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s103, v77, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s102, v77, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s101, v77, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s100, v77, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s99, v77, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s98, v77, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s97, v77, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s96, v77, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s87, v76, 31
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s86, v76, 30
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s85, v76, 29
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s84, v76, 28
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s83, v76, 27
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s82, v76, 26
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s81, v76, 25
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s80, v76, 24
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s71, v76, 23
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s70, v76, 22
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s69, v76, 21
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s68, v76, 20
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v76, 19
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v76, 18
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v76, 17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v76, 16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v76, 15
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v76, 14
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v76, 13
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v76, 12
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v76, 11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v76, 10
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v76, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v76, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v76, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v76, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v76, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v76, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v76, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v76, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v76, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v76, 0
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:88
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:92
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v16f64_to_v128i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:80
+; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:84
+; GFX11-NEXT:    scratch_store_b32 off, v78, s32 offset:88
+; GFX11-NEXT:    scratch_store_b32 off, v79, s32 offset:92
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v76, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v77, s96, 0
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v2
+; GFX11-NEXT:    v_writelane_b32 v76, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v77, s97, 1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v5
+; GFX11-NEXT:    v_writelane_b32 v76, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v77, s98, 2
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v8
+; GFX11-NEXT:    v_writelane_b32 v76, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v77, s99, 3
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v11
+; GFX11-NEXT:    v_writelane_b32 v76, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v77, s100, 4
+; GFX11-NEXT:    v_readfirstlane_b32 s15, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s40, v13
+; GFX11-NEXT:    v_readfirstlane_b32 s41, v14
+; GFX11-NEXT:    v_writelane_b32 v76, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v77, s101, 5
+; GFX11-NEXT:    s_mov_b32 vcc_hi, 0
+; GFX11-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
+; GFX11-NEXT:    s_clause 0x13
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:76
+; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:72
+; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:68
+; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:64
+; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:60
+; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:56
+; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:52
+; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:48
+; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:44
+; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:40
+; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:36
+; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:32
+; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:28
+; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:24
+; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:20
+; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:16
+; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:12
+; GFX11-NEXT:    scratch_store_b32 off, v73, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v74, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v75, s32
+; GFX11-NEXT:    v_writelane_b32 v76, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v77, s102, 6
+; GFX11-NEXT:    ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
+; GFX11-NEXT:    ; implicit-def: $vgpr79 : SGPR spill to VGPR lane
+; GFX11-NEXT:    v_writelane_b32 v76, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v77, s103, 7
+; GFX11-NEXT:    v_writelane_b32 v76, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s104, 8
+; GFX11-NEXT:    v_writelane_b32 v76, s49, 9
+; GFX11-NEXT:    v_writelane_b32 v76, s50, 10
+; GFX11-NEXT:    v_writelane_b32 v76, s51, 11
+; GFX11-NEXT:    v_writelane_b32 v76, s52, 12
+; GFX11-NEXT:    v_writelane_b32 v76, s53, 13
+; GFX11-NEXT:    v_writelane_b32 v76, s54, 14
+; GFX11-NEXT:    v_writelane_b32 v76, s55, 15
+; GFX11-NEXT:    v_writelane_b32 v76, s64, 16
+; GFX11-NEXT:    v_writelane_b32 v76, s65, 17
+; GFX11-NEXT:    v_writelane_b32 v76, s66, 18
+; GFX11-NEXT:    v_writelane_b32 v76, s67, 19
+; GFX11-NEXT:    v_writelane_b32 v76, s68, 20
+; GFX11-NEXT:    v_writelane_b32 v76, s69, 21
+; GFX11-NEXT:    v_writelane_b32 v76, s70, 22
+; GFX11-NEXT:    v_writelane_b32 v76, s71, 23
+; GFX11-NEXT:    v_writelane_b32 v76, s80, 24
+; GFX11-NEXT:    v_writelane_b32 v76, s81, 25
+; GFX11-NEXT:    v_writelane_b32 v76, s82, 26
+; GFX11-NEXT:    v_writelane_b32 v76, s83, 27
+; GFX11-NEXT:    v_writelane_b32 v76, s84, 28
+; GFX11-NEXT:    v_writelane_b32 v76, s85, 29
+; GFX11-NEXT:    v_writelane_b32 v76, s86, 30
+; GFX11-NEXT:    v_writelane_b32 v76, s87, 31
+; GFX11-NEXT:    s_cbranch_scc0 .LBB73_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s50, s41, 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 0
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 8
+; GFX11-NEXT:    s_lshr_b32 s49, s41, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s41, 8
+; GFX11-NEXT:    s_lshr_b32 s52, s40, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 1
+; GFX11-NEXT:    s_lshr_b32 s42, s11, 24
+; GFX11-NEXT:    s_lshr_b32 s51, s40, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s15, 24
+; GFX11-NEXT:    s_lshr_b32 s38, s15, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 2
+; GFX11-NEXT:    s_lshr_b32 s42, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s37, s15, 8
+; GFX11-NEXT:    s_lshr_b32 s54, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s53, s14, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s11, 8
+; GFX11-NEXT:    s_lshr_b32 s36, s13, 24
+; GFX11-NEXT:    s_lshr_b32 s64, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s55, s12, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 4
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 24
+; GFX11-NEXT:    s_lshr_b32 s66, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s65, s10, 8
+; GFX11-NEXT:    s_lshr_b32 s68, s8, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 5
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s67, s8, 8
+; GFX11-NEXT:    s_lshr_b32 s70, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s69, s6, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 6
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 8
+; GFX11-NEXT:    s_lshr_b32 s80, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s71, s4, 8
+; GFX11-NEXT:    s_lshr_b32 s82, s28, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 7
+; GFX11-NEXT:    s_lshr_b32 s42, s7, 24
+; GFX11-NEXT:    s_lshr_b32 s81, s28, 8
+; GFX11-NEXT:    s_lshr_b32 s84, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s83, s26, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s86, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s85, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s96, s22, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 9
+; GFX11-NEXT:    s_lshr_b32 s42, s7, 8
+; GFX11-NEXT:    s_lshr_b32 s87, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s98, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s97, s20, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 10
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 24
+; GFX11-NEXT:    s_lshr_b32 s100, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s99, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s102, s16, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 11
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s101, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s104, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s103, s2, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 12
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 8
+; GFX11-NEXT:    s_lshr_b32 s35, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s34, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 13
+; GFX11-NEXT:    s_lshr_b32 s42, s29, 24
+; GFX11-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
+; GFX11-NEXT:    s_lshr_b64 s[74:75], s[12:13], 24
+; GFX11-NEXT:    s_lshr_b64 s[76:77], s[10:11], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 14
+; GFX11-NEXT:    s_lshr_b32 s42, s29, 16
+; GFX11-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
+; GFX11-NEXT:    s_lshr_b64 s[88:89], s[6:7], 24
+; GFX11-NEXT:    s_lshr_b64 s[90:91], s[4:5], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 15
+; GFX11-NEXT:    s_lshr_b32 s42, s29, 8
+; GFX11-NEXT:    s_lshr_b64 s[92:93], s[28:29], 24
+; GFX11-NEXT:    s_lshr_b64 s[94:95], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
+; GFX11-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 17
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 18
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 19
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 20
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 21
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 22
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 23
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 25
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 26
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 27
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 28
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 29
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 30
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 31
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 24
+; GFX11-NEXT:    v_writelane_b32 v79, s42, 0
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v79, s42, 1
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
+; GFX11-NEXT:    v_writelane_b32 v79, s42, 2
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v79, s42, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    v_writelane_b32 v79, s42, 4
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v79, s42, 5
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
+; GFX11-NEXT:    v_writelane_b32 v79, s42, 6
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v79, s42, 7
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 8
+; GFX11-NEXT:    v_writelane_b32 v79, s42, 8
+; GFX11-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
+; GFX11-NEXT:    s_cbranch_vccnz .LBB73_4
+; GFX11-NEXT:  .LBB73_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[23:24], s[24:25], 1.0
+; GFX11-NEXT:    v_add_f64 v[28:29], s[22:23], 1.0
+; GFX11-NEXT:    v_add_f64 v[32:33], s[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[5:6], s[12:13], 1.0
+; GFX11-NEXT:    v_add_f64 v[36:37], s[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[52:53], s[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[1:2], s[40:41], 1.0
+; GFX11-NEXT:    v_add_f64 v[3:4], s[14:15], 1.0
+; GFX11-NEXT:    v_add_f64 v[7:8], s[10:11], 1.0
+; GFX11-NEXT:    v_add_f64 v[9:10], s[8:9], 1.0
+; GFX11-NEXT:    v_add_f64 v[11:12], s[6:7], 1.0
+; GFX11-NEXT:    v_add_f64 v[13:14], s[4:5], 1.0
+; GFX11-NEXT:    v_add_f64 v[15:16], s[28:29], 1.0
+; GFX11-NEXT:    v_add_f64 v[19:20], s[26:27], 1.0
+; GFX11-NEXT:    v_add_f64 v[48:49], s[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[64:65], s[0:1], 1.0
+; GFX11-NEXT:    v_lshrrev_b64 v[66:67], 24, v[23:24]
+; GFX11-NEXT:    v_lshrrev_b64 v[67:68], 24, v[28:29]
+; GFX11-NEXT:    v_lshrrev_b64 v[68:69], 24, v[32:33]
+; GFX11-NEXT:    v_lshrrev_b64 v[25:26], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b64 v[69:70], 24, v[36:37]
+; GFX11-NEXT:    v_lshrrev_b64 v[80:81], 24, v[52:53]
+; GFX11-NEXT:    v_lshrrev_b64 v[17:18], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[21:22], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[26:27], 24, v[7:8]
+; GFX11-NEXT:    v_lshrrev_b64 v[30:31], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[34:35], 24, v[11:12]
+; GFX11-NEXT:    v_lshrrev_b64 v[38:39], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
+; GFX11-NEXT:    v_lshrrev_b64 v[54:55], 24, v[19:20]
+; GFX11-NEXT:    v_lshrrev_b64 v[70:71], 24, v[48:49]
+; GFX11-NEXT:    v_lshrrev_b64 v[81:82], 24, v[64:65]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 24, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 24, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 8, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v99, 24, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v100, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v101, 8, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v97, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v98, 8, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v112, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v114, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v116, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v102, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v103, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v119, 24, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v128, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v129, 8, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v113, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v115, 8, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v132, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v133, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v134, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v117, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v118, 8, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v145, 24, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v146, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v147, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v130, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v131, 8, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v150, 24, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v151, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v162, 8, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v135, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v144, 8, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v164, 24, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v166, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v167, 8, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v148, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v149, 8, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v178, 24, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v179, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v180, 8, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v160, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v161, 8, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v183, 24, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v40, 16, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v42, 8, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v163, 16, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v165, 8, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v45, 24, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v47, 16, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v56, 8, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v176, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v177, 8, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v57, 24, v49
+; GFX11-NEXT:    v_lshrrev_b32_e32 v58, 16, v49
+; GFX11-NEXT:    v_lshrrev_b32_e32 v59, 8, v49
+; GFX11-NEXT:    v_lshrrev_b32_e32 v181, 16, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v182, 8, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v60, 24, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v61, 16, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v62, 8, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v41, 16, v52
+; GFX11-NEXT:    v_lshrrev_b32_e32 v43, 8, v52
+; GFX11-NEXT:    v_lshrrev_b32_e32 v63, 24, v65
+; GFX11-NEXT:    v_lshrrev_b32_e32 v72, 16, v65
+; GFX11-NEXT:    v_lshrrev_b32_e32 v73, 8, v65
+; GFX11-NEXT:    v_lshrrev_b32_e32 v44, 16, v64
+; GFX11-NEXT:    v_lshrrev_b32_e32 v46, 8, v64
+; GFX11-NEXT:    s_branch .LBB73_5
+; GFX11-NEXT:  .LBB73_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr103
+; GFX11-NEXT:    ; implicit-def: $sgpr104
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr101
+; GFX11-NEXT:    ; implicit-def: $sgpr102
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr99
+; GFX11-NEXT:    ; implicit-def: $sgpr100
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr97
+; GFX11-NEXT:    ; implicit-def: $sgpr98
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr87
+; GFX11-NEXT:    ; implicit-def: $sgpr96
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr85
+; GFX11-NEXT:    ; implicit-def: $sgpr86
+; GFX11-NEXT:    ; implicit-def: $sgpr83
+; GFX11-NEXT:    ; implicit-def: $sgpr84
+; GFX11-NEXT:    ; implicit-def: $sgpr81
+; GFX11-NEXT:    ; implicit-def: $sgpr82
+; GFX11-NEXT:    ; implicit-def: $sgpr71
+; GFX11-NEXT:    ; implicit-def: $sgpr80
+; GFX11-NEXT:    ; implicit-def: $sgpr69
+; GFX11-NEXT:    ; implicit-def: $sgpr70
+; GFX11-NEXT:    ; implicit-def: $sgpr67
+; GFX11-NEXT:    ; implicit-def: $sgpr68
+; GFX11-NEXT:    ; implicit-def: $sgpr65
+; GFX11-NEXT:    ; implicit-def: $sgpr66
+; GFX11-NEXT:    ; implicit-def: $sgpr55
+; GFX11-NEXT:    ; implicit-def: $sgpr64
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr53
+; GFX11-NEXT:    ; implicit-def: $sgpr54
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr51
+; GFX11-NEXT:    ; implicit-def: $sgpr52
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr50
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; kill: killed $sgpr43
+; GFX11-NEXT:    s_branch .LBB73_2
+; GFX11-NEXT:  .LBB73_4:
+; GFX11-NEXT:    v_dual_mov_b32 v64, s0 :: v_dual_mov_b32 v65, s1
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 0
+; GFX11-NEXT:    v_dual_mov_b32 v1, s40 :: v_dual_mov_b32 v2, s41
+; GFX11-NEXT:    v_dual_mov_b32 v3, s14 :: v_dual_mov_b32 v4, s15
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_mov_b32_e32 v87, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 1
+; GFX11-NEXT:    v_mov_b32_e32 v39, s54
+; GFX11-NEXT:    v_dual_mov_b32 v5, s12 :: v_dual_mov_b32 v6, s13
+; GFX11-NEXT:    v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
+; GFX11-NEXT:    v_mov_b32_e32 v96, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 2
+; GFX11-NEXT:    v_dual_mov_b32 v9, s8 :: v_dual_mov_b32 v10, s9
+; GFX11-NEXT:    v_dual_mov_b32 v11, s6 :: v_dual_mov_b32 v12, s7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_mov_b32_e32 v99, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 3
+; GFX11-NEXT:    v_mov_b32_e32 v55, s53
+; GFX11-NEXT:    v_dual_mov_b32 v13, s4 :: v_dual_mov_b32 v14, s5
+; GFX11-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    v_mov_b32_e32 v100, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 4
+; GFX11-NEXT:    v_dual_mov_b32 v19, s26 :: v_dual_mov_b32 v20, s27
+; GFX11-NEXT:    v_dual_mov_b32 v23, s24 :: v_dual_mov_b32 v24, s25
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v101, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 5
+; GFX11-NEXT:    v_dual_mov_b32 v27, s51 :: v_dual_mov_b32 v28, s22
+; GFX11-NEXT:    v_dual_mov_b32 v29, s23 :: v_dual_mov_b32 v32, s20
+; GFX11-NEXT:    v_dual_mov_b32 v33, s21 :: v_dual_mov_b32 v112, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 6
+; GFX11-NEXT:    v_dual_mov_b32 v31, s49 :: v_dual_mov_b32 v36, s18
+; GFX11-NEXT:    v_dual_mov_b32 v37, s19 :: v_dual_mov_b32 v48, s16
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v49, s17 :: v_dual_mov_b32 v114, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 7
+; GFX11-NEXT:    v_dual_mov_b32 v35, s48 :: v_dual_mov_b32 v52, s2
+; GFX11-NEXT:    v_dual_mov_b32 v53, s3 :: v_dual_mov_b32 v44, s35
+; GFX11-NEXT:    v_dual_mov_b32 v41, s104 :: v_dual_mov_b32 v116, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 8
+; GFX11-NEXT:    v_dual_mov_b32 v46, s34 :: v_dual_mov_b32 v43, s103
+; GFX11-NEXT:    v_dual_mov_b32 v181, s102 :: v_dual_mov_b32 v182, s101
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_mov_b32_e32 v119, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 9
+; GFX11-NEXT:    v_dual_mov_b32 v51, s39 :: v_dual_mov_b32 v176, s100
+; GFX11-NEXT:    v_mov_b32_e32 v177, s99
+; GFX11-NEXT:    v_dual_mov_b32 v163, s98 :: v_dual_mov_b32 v160, s96
+; GFX11-NEXT:    v_mov_b32_e32 v128, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 10
+; GFX11-NEXT:    v_dual_mov_b32 v165, s97 :: v_dual_mov_b32 v148, s86
+; GFX11-NEXT:    v_dual_mov_b32 v161, s87 :: v_dual_mov_b32 v144, s83
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_mov_b32_e32 v129, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 11
+; GFX11-NEXT:    v_mov_b32_e32 v71, s38
+; GFX11-NEXT:    v_dual_mov_b32 v149, s85 :: v_dual_mov_b32 v130, s82
+; GFX11-NEXT:    v_dual_mov_b32 v135, s84 :: v_dual_mov_b32 v118, s71
+; GFX11-NEXT:    v_mov_b32_e32 v132, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 12
+; GFX11-NEXT:    v_dual_mov_b32 v131, s81 :: v_dual_mov_b32 v102, s68
+; GFX11-NEXT:    v_dual_mov_b32 v117, s80 :: v_dual_mov_b32 v98, s65
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_mov_b32_e32 v133, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 13
+; GFX11-NEXT:    v_mov_b32_e32 v83, s37
+; GFX11-NEXT:    v_dual_mov_b32 v113, s70 :: v_dual_mov_b32 v84, s64
+; GFX11-NEXT:    v_dual_mov_b32 v115, s69 :: v_dual_mov_b32 v86, s55
+; GFX11-NEXT:    v_mov_b32_e32 v134, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 14
+; GFX11-NEXT:    v_dual_mov_b32 v103, s67 :: v_dual_mov_b32 v18, s52
+; GFX11-NEXT:    v_dual_mov_b32 v97, s66 :: v_dual_mov_b32 v22, s50
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_mov_b32_e32 v145, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 15
+; GFX11-NEXT:    v_mov_b32_e32 v85, s36
+; GFX11-NEXT:    v_dual_mov_b32 v81, s42 :: v_dual_mov_b32 v38, s90
+; GFX11-NEXT:    v_dual_mov_b32 v69, s56 :: v_dual_mov_b32 v34, s88
+; GFX11-NEXT:    v_mov_b32_e32 v146, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 16
+; GFX11-NEXT:    v_dual_mov_b32 v67, s60 :: v_dual_mov_b32 v30, s78
+; GFX11-NEXT:    v_dual_mov_b32 v26, s76 :: v_dual_mov_b32 v25, s74
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_mov_b32_e32 v147, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 17
+; GFX11-NEXT:    v_mov_b32_e32 v21, s72
+; GFX11-NEXT:    v_dual_mov_b32 v17, s62 :: v_dual_mov_b32 v80, s44
+; GFX11-NEXT:    v_mov_b32_e32 v70, s46
+; GFX11-NEXT:    v_mov_b32_e32 v150, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 18
+; GFX11-NEXT:    v_mov_b32_e32 v68, s58
+; GFX11-NEXT:    v_mov_b32_e32 v66, s30
+; GFX11-NEXT:    v_mov_b32_e32 v54, s94
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_dual_mov_b32 v50, s92 :: v_dual_mov_b32 v151, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 19
+; GFX11-NEXT:    v_mov_b32_e32 v162, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 20
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v164, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 21
+; GFX11-NEXT:    v_mov_b32_e32 v166, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 22
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v167, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 23
+; GFX11-NEXT:    v_mov_b32_e32 v178, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 24
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v179, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 25
+; GFX11-NEXT:    v_mov_b32_e32 v180, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 26
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v183, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 27
+; GFX11-NEXT:    v_mov_b32_e32 v40, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 28
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v42, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 29
+; GFX11-NEXT:    v_mov_b32_e32 v45, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 30
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v47, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 31
+; GFX11-NEXT:    v_mov_b32_e32 v56, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v79, 0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v57, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v79, 1
+; GFX11-NEXT:    v_mov_b32_e32 v58, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v79, 2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v59, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v79, 3
+; GFX11-NEXT:    v_mov_b32_e32 v60, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v79, 4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v61, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v79, 5
+; GFX11-NEXT:    v_mov_b32_e32 v62, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v79, 6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v63, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v79, 7
+; GFX11-NEXT:    v_mov_b32_e32 v72, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v79, 8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v73, s0
+; GFX11-NEXT:  .LBB73_5: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v82, 8, v46
+; GFX11-NEXT:    v_and_b32_e32 v64, 0xff, v64
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v81
+; GFX11-NEXT:    v_and_b32_e32 v65, 0xff, v65
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v80
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xff, v52
+; GFX11-NEXT:    v_or_b32_e32 v64, v64, v82
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xff, v44
+; GFX11-NEXT:    v_lshlrev_b32_e32 v44, 8, v63
+; GFX11-NEXT:    v_lshlrev_b32_e32 v43, 8, v43
+; GFX11-NEXT:    v_and_b32_e32 v41, 0xff, v41
+; GFX11-NEXT:    v_and_b32_e32 v64, 0xffff, v64
+; GFX11-NEXT:    v_or_b32_e32 v81, v82, v81
+; GFX11-NEXT:    v_lshlrev_b32_e32 v82, 8, v73
+; GFX11-NEXT:    v_or_b32_e32 v52, v52, v43
+; GFX11-NEXT:    v_or_b32_e32 v80, v41, v80
+; GFX11-NEXT:    v_and_b32_e32 v53, 0xff, v53
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 16, v81
+; GFX11-NEXT:    v_or_b32_e32 v65, v65, v82
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xff, v72
+; GFX11-NEXT:    v_lshlrev_b32_e32 v41, 8, v60
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v52
+; GFX11-NEXT:    v_or_b32_e32 v72, v64, v81
+; GFX11-NEXT:    v_and_b32_e32 v64, 0xffff, v65
+; GFX11-NEXT:    v_or_b32_e32 v82, v82, v44
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v62
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xff, v48
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
+; GFX11-NEXT:    v_and_b32_e32 v49, 0xff, v49
+; GFX11-NEXT:    v_lshlrev_b32_e32 v65, 16, v82
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xff, v61
+; GFX11-NEXT:    v_or_b32_e32 v53, v53, v81
+; GFX11-NEXT:    v_and_b32_e32 v81, 0xff, v181
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xff, v36
+; GFX11-NEXT:    v_or_b32_e32 v73, v64, v65
+; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 16, v80
+; GFX11-NEXT:    v_or_b32_e32 v65, v82, v41
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v182
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v69
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xff, v37
+; GFX11-NEXT:    v_or_b32_e32 v74, v52, v64
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v53
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 16, v65
+; GFX11-NEXT:    v_or_b32_e32 v48, v48, v80
+; GFX11-NEXT:    v_or_b32_e32 v64, v81, v70
+; GFX11-NEXT:    v_lshlrev_b32_e32 v65, 8, v59
+; GFX11-NEXT:    v_and_b32_e32 v70, 0xff, v58
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v57
+; GFX11-NEXT:    v_or_b32_e32 v75, v52, v53
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v48
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 16, v64
+; GFX11-NEXT:    v_or_b32_e32 v49, v49, v65
+; GFX11-NEXT:    v_or_b32_e32 v53, v70, v80
+; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 8, v177
+; GFX11-NEXT:    v_and_b32_e32 v65, 0xff, v176
+; GFX11-NEXT:    v_or_b32_e32 v43, v48, v52
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v49
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 16, v53
+; GFX11-NEXT:    v_or_b32_e32 v36, v36, v64
+; GFX11-NEXT:    v_or_b32_e32 v52, v65, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 8, v56
+; GFX11-NEXT:    v_and_b32_e32 v64, 0xff, v47
+; GFX11-NEXT:    v_lshlrev_b32_e32 v65, 8, v45
+; GFX11-NEXT:    v_or_b32_e32 v44, v48, v49
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v52
+; GFX11-NEXT:    v_or_b32_e32 v37, v37, v53
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v32
+; GFX11-NEXT:    v_or_b32_e32 v49, v64, v65
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 8, v165
+; GFX11-NEXT:    v_and_b32_e32 v53, 0xff, v163
+; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 8, v68
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xff, v33
+; GFX11-NEXT:    v_lshlrev_b32_e32 v65, 8, v42
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
+; GFX11-NEXT:    v_or_b32_e32 v32, v32, v52
+; GFX11-NEXT:    v_or_b32_e32 v52, v53, v64
+; GFX11-NEXT:    v_or_b32_e32 v33, v33, v65
+; GFX11-NEXT:    v_or_b32_e32 v45, v36, v48
+; GFX11-NEXT:    v_or_b32_e32 v46, v37, v49
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xff, v40
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v52
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v183
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 8, v161
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xff, v160
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 8, v67
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 8, v180
+; GFX11-NEXT:    v_and_b32_e32 v65, 0xff, v179
+; GFX11-NEXT:    v_lshlrev_b32_e32 v67, 8, v178
+; GFX11-NEXT:    v_or_b32_e32 v37, v37, v48
+; GFX11-NEXT:    v_or_b32_e32 v28, v28, v49
+; GFX11-NEXT:    v_or_b32_e32 v48, v52, v53
+; GFX11-NEXT:    v_or_b32_e32 v29, v29, v64
+; GFX11-NEXT:    v_or_b32_e32 v49, v65, v67
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
+; GFX11-NEXT:    v_or_b32_e32 v67, v32, v36
+; GFX11-NEXT:    v_or_b32_e32 v68, v33, v37
+; GFX11-NEXT:    v_or_b32_e32 v69, v28, v48
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xff, v23
+; GFX11-NEXT:    v_or_b32_e32 v70, v29, v49
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v149
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v148
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v66
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v167
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xff, v166
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 8, v164
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v144
+; GFX11-NEXT:    v_or_b32_e32 v23, v23, v28
+; GFX11-NEXT:    v_or_b32_e32 v28, v29, v32
+; GFX11-NEXT:    v_or_b32_e32 v24, v24, v33
+; GFX11-NEXT:    v_or_b32_e32 v29, v36, v37
+; GFX11-NEXT:    v_or_b32_e32 v19, v19, v48
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v135
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v54
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 8, v162
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xff, v151
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v150
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 8, v131
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xff, v130
+; GFX11-NEXT:    v_lshlrev_b32_e32 v50, 8, v50
+; GFX11-NEXT:    v_or_b32_e32 v32, v32, v33
+; GFX11-NEXT:    v_or_b32_e32 v20, v20, v36
+; GFX11-NEXT:    v_or_b32_e32 v33, v37, v48
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v49
+; GFX11-NEXT:    v_or_b32_e32 v36, v52, v50
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
+; GFX11-NEXT:    v_or_b32_e32 v148, v23, v28
+; GFX11-NEXT:    v_or_b32_e32 v150, v19, v32
+; GFX11-NEXT:    v_or_b32_e32 v151, v20, v33
+; GFX11-NEXT:    v_or_b32_e32 v130, v15, v36
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v147
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v146
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v145
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 8, v118
+; GFX11-NEXT:    v_or_b32_e32 v149, v24, v29
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v117
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v38
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 8, v134
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v16, v19, v20
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v23
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v133
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 8, v132
+; GFX11-NEXT:    v_or_b32_e32 v19, v24, v28
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v29
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 8, v115
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v113
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 8, v34
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v129
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xff, v128
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 8, v119
+; GFX11-NEXT:    v_or_b32_e32 v20, v20, v23
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v24
+; GFX11-NEXT:    v_or_b32_e32 v23, v28, v29
+; GFX11-NEXT:    v_or_b32_e32 v12, v12, v32
+; GFX11-NEXT:    v_or_b32_e32 v24, v33, v34
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
+; GFX11-NEXT:    v_or_b32_e32 v131, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v132, v13, v19
+; GFX11-NEXT:    v_or_b32_e32 v133, v14, v20
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v103
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v102
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v30
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v116
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v114
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v112
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v23
+; GFX11-NEXT:    v_or_b32_e32 v12, v12, v24
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 8, v98
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v13
+; GFX11-NEXT:    v_or_b32_e32 v13, v14, v15
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v16
+; GFX11-NEXT:    v_or_b32_e32 v14, v19, v20
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v97
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v26
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v101
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 8, v86
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xff, v84
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v23
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v100
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 8, v99
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v19
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v24
+; GFX11-NEXT:    v_or_b32_e32 v19, v26, v25
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_or_b32_e32 v16, v20, v23
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_or_b32_e32 v13, v9, v13
+; GFX11-NEXT:    v_or_b32_e32 v14, v10, v14
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v15
+; GFX11-NEXT:    v_or_b32_e32 v9, v5, v19
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 8, v96
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v87
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v85
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v39
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v21
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v16
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v55
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v83
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v10, v15
+; GFX11-NEXT:    v_or_b32_e32 v10, v19, v20
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v16
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v51
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v71
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v27
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xff, v18
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v35
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v31
+; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 8, v22
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v16
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v19
+; GFX11-NEXT:    v_or_b32_e32 v16, v18, v17
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v20
+; GFX11-NEXT:    v_or_b32_e32 v17, v21, v22
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-NEXT:    v_or_b32_e32 v10, v5, v6
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    scratch_store_b128 v0, v[72:75], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[43:46], off offset:16
+; GFX11-NEXT:    v_or_b32_e32 v1, v3, v15
+; GFX11-NEXT:    v_or_b32_e32 v2, v4, v18
+; GFX11-NEXT:    v_or_b32_e32 v3, v19, v16
+; GFX11-NEXT:    v_or_b32_e32 v4, v20, v17
+; GFX11-NEXT:    s_clause 0x5
+; GFX11-NEXT:    scratch_store_b128 v0, v[67:70], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[148:151], off offset:48
+; GFX11-NEXT:    scratch_store_b128 v0, v[130:133], off offset:64
+; GFX11-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
+; GFX11-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
+; GFX11-NEXT:    s_clause 0x13
+; GFX11-NEXT:    scratch_load_b32 v75, off, s32
+; GFX11-NEXT:    scratch_load_b32 v74, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v73, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:12
+; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:16
+; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:20
+; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:24
+; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:28
+; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:32
+; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:36
+; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:40
+; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:44
+; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:48
+; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:52
+; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:56
+; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:60
+; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:64
+; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:68
+; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:72
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:76
+; GFX11-NEXT:    v_readlane_b32 s104, v77, 8
+; GFX11-NEXT:    v_readlane_b32 s103, v77, 7
+; GFX11-NEXT:    v_readlane_b32 s102, v77, 6
+; GFX11-NEXT:    v_readlane_b32 s101, v77, 5
+; GFX11-NEXT:    v_readlane_b32 s100, v77, 4
+; GFX11-NEXT:    v_readlane_b32 s99, v77, 3
+; GFX11-NEXT:    v_readlane_b32 s98, v77, 2
+; GFX11-NEXT:    v_readlane_b32 s97, v77, 1
+; GFX11-NEXT:    v_readlane_b32 s96, v77, 0
+; GFX11-NEXT:    v_readlane_b32 s87, v76, 31
+; GFX11-NEXT:    v_readlane_b32 s86, v76, 30
+; GFX11-NEXT:    v_readlane_b32 s85, v76, 29
+; GFX11-NEXT:    v_readlane_b32 s84, v76, 28
+; GFX11-NEXT:    v_readlane_b32 s83, v76, 27
+; GFX11-NEXT:    v_readlane_b32 s82, v76, 26
+; GFX11-NEXT:    v_readlane_b32 s81, v76, 25
+; GFX11-NEXT:    v_readlane_b32 s80, v76, 24
+; GFX11-NEXT:    v_readlane_b32 s71, v76, 23
+; GFX11-NEXT:    v_readlane_b32 s70, v76, 22
+; GFX11-NEXT:    v_readlane_b32 s69, v76, 21
+; GFX11-NEXT:    v_readlane_b32 s68, v76, 20
+; GFX11-NEXT:    v_readlane_b32 s67, v76, 19
+; GFX11-NEXT:    v_readlane_b32 s66, v76, 18
+; GFX11-NEXT:    v_readlane_b32 s65, v76, 17
+; GFX11-NEXT:    v_readlane_b32 s64, v76, 16
+; GFX11-NEXT:    v_readlane_b32 s55, v76, 15
+; GFX11-NEXT:    v_readlane_b32 s54, v76, 14
+; GFX11-NEXT:    v_readlane_b32 s53, v76, 13
+; GFX11-NEXT:    v_readlane_b32 s52, v76, 12
+; GFX11-NEXT:    v_readlane_b32 s51, v76, 11
+; GFX11-NEXT:    v_readlane_b32 s50, v76, 10
+; GFX11-NEXT:    v_readlane_b32 s49, v76, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v76, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v76, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v76, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v76, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v76, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v76, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v76, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v76, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v76, 0
+; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:80
+; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:84
+; GFX11-NEXT:    scratch_load_b32 v78, off, s32 offset:88
+; GFX11-NEXT:    scratch_load_b32 v79, off, s32 offset:92
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -133785,462 +126993,482 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v16f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:436
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:320
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v86, v0
+; GFX11-TRUE16-NEXT:    s_clause 0x1f
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:476
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:472
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:468
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:464
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:460
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:456
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:452
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:448
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:444
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:352
+; GFX11-TRUE16-NEXT:    s_clause 0x7
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:320
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v26 :: v_dual_mov_b32 v35, v24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v22 :: v_dual_mov_b32 v37, v20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v18 :: v_dual_mov_b32 v39, v16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v14 :: v_dual_mov_b32 v49, v12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v10 :: v_dual_mov_b32 v51, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v6 :: v_dual_mov_b32 v53, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, v2 :: v_dual_mov_b32 v55, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
 ; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:316
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:312
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:308
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:304
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:300
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:296
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:292
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v2, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v16, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v18, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v20, off, s32 offset:72
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v22, off, s32 offset:80
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v24, off, s32 offset:88
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v26, off, s32 offset:96
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v28, off, s32 offset:104
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v30, off, s32 offset:112
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:120
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:128
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:136
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:144
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:152
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:160
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:168
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:176
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:184
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:200
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:208
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:216
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:224
 ; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v94, off, s32 offset:240
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v95, off, s32 offset:248
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v104, off, s32 offset:256
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v105, off, s32 offset:264
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v106, off, s32 offset:272
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v107, off, s32 offset:280
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v108, off, s32 offset:288
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v109, off, s32 offset:296
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v110, off, s32 offset:304
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v111, off, s32 offset:312
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:308
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:300
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:292
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:284
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:276
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:268
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:260
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:252
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:244
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:228
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:220
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:212
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:204
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:196
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:188
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:180
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:172
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:164
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:156
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:148
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:140
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:132
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v89, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v88, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v90, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v91, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v93, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v92, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v30 :: v_dual_mov_b32 v33, v29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v28 :: v_dual_mov_b32 v34, v27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v26 :: v_dual_mov_b32 v37, v25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v24 :: v_dual_mov_b32 v38, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v22 :: v_dual_mov_b32 v49, v21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v20 :: v_dual_mov_b32 v51, v19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v18 :: v_dual_mov_b32 v53, v17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v55, v16 :: v_dual_mov_b32 v54, v15
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, v14 :: v_dual_mov_b32 v65, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v67, v12 :: v_dual_mov_b32 v66, v11
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v68, v10 :: v_dual_mov_b32 v69, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v71, v8 :: v_dual_mov_b32 v70, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v80, v6 :: v_dual_mov_b32 v81, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v83, v4 :: v_dual_mov_b32 v82, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v84, v2 :: v_dual_mov_b32 v85, v1
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:124
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:116
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:108
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:100
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:92
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:84
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:76
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:68
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v89, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v90, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v91, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v92, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v93, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v76, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v77, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v78, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v79, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v88, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v63, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v72, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v73, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v74, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v75, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v57, 8, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v58, 8, v4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v59, 8, v6
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v60, 8, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v61, 8, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v40, 8, v12
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 8, v14
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v42, 8, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v43, 8, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v44, 8, v20
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v167, 8, v22
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v176, 8, v24
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v177, 8, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v178, 8, v28
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v179, 8, v30
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v149, 8, v31
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v150, 8, v87
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(61)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v151, 8, v96
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(60)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v160, 8, v97
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(59)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v161, 8, v98
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(58)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v132, 8, v99
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(57)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v133, 8, v113
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(56)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v134, 8, v114
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(55)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v135, 8, v115
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(54)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v144, 8, v116
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(53)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v119, 8, v117
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(52)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v128, 8, v128
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(51)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v129, 8, v129
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(50)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v130, 8, v130
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(49)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v131, 8, v131
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(48)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v113, 8, v94
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(47)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v114, 8, v95
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(46)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v115, 8, v104
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(45)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v116, 8, v105
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(44)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v117, 8, v106
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(43)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v107
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(42)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v108
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(41)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v109
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(40)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v98, 8, v110
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(39)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v99, 8, v111
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB75_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v71
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v53
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v90
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v91
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v68
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v65
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v67
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v53
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s26, 0xff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v49
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s27, 8
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v76
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v77
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v64
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v78
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v79
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v49
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v51
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v37
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v88
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v63
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v37
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v72
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v73
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v33
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v33
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v74
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v75
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v91
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v0, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v32
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v92
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v93
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v62
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v57
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v58
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v79
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v90
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v88
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v89
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v56
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v47
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v59
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v60
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v75
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v78
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v77
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v76
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v45
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v61
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v40
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v63
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v74
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v72
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v73
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v183
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v182
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v41
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v42
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v59
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v62
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v60
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v61
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v43
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v44
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v47
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v56
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v57
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v165
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v167
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v176
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v43
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v44
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v45
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v164
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v163
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v177
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v178
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v183
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v42
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v41
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v40
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v148
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v179
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v149
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v179
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v182
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v147
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v146
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v150
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v151
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v167
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v178
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v176
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v177
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v118
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v160
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v161
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v163
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v164
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v165
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v103
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v132
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v133
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v151
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v160
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v161
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v102
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v101
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v134
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v135
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v147
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v150
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v149
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v148
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v100
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v86
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v144
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v119
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v135
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v146
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v144
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v85
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v84
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v128
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v129
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v131
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v134
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v132
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v133
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v83
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v130
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v131
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v119
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v130
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v128
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v129
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v81
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v80
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v113
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v114
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v114
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v118
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v116
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v117
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v115
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v116
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v102
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v115
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v113
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v68
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v117
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v87
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v98
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v103
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v100
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v67
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v96
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v97
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v87
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v99
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v96
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v97
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v64
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v98
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v99
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v86
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v85
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v89
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
@@ -134273,8 +127501,18 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s10, 16
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v51
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v93
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v92
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB75_3
@@ -134292,445 +127530,401 @@ define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a,
 ; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s4
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
 ; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v86
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v55
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v54
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v52
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v85
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v51
 ; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v84
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v81
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v82
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v38
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v89, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v90, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v91, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v92, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v93, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v69
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v88, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v74, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v75, v12
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v68
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v65
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v64
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v8, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v54
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v49
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v48
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v39
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v37
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v8, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v46
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v181
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v76, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v77, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v78, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v79, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v63, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v72, v8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v73, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v49
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v35
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v9, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v33
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v3, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v93
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v61, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v43, v16
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v44, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v14, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v32
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v92
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v62
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v56
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v47
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v45
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v183
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v182
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v91
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v90
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v89
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v12, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v13, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v88
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v79
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v78
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v162
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v145
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v118
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v57, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v58, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v59, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v60, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v40, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v41, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v42, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v12, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v13, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v77
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v76
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v75
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v74
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v73
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v14, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v72
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v63
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v62
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v61
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v14, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v15, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v59
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v17, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v15
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v58
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v3, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v57
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v56
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v179, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v160, v21
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v161, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v18
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v19, v15
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v166
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v165
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v164
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v163
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v148
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v147
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v146
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v47
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v46
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v45
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v17, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v18, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v44
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v43
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v42
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v100
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v83
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v167, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v176, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v177, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v178, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v149, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v150, v18
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v151, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v17, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v18, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v41
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v40
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v183
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v182
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v181
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v19, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v180
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v178
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v177
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v19, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v20, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v176
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v167
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v22, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v166
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v3, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v165
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v164
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v144, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v130, v26
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v131, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v19, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v24, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v112
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v103
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v102
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v86
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v85
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v163
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v162
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v161
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v22, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v23, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v160
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v151
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v150
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v26, v27
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v69
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v132, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v133, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v134, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v135, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v119, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v128, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v129, v25
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v22, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v23, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v149
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v148
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v147
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v146
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v145
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v24, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v144
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v135
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v134
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v133
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v24, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v25, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v132
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v131
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v27, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v130
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v3, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v128
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v129
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v119
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v113
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v27, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v115
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v117, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v24, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v29, v25
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v81
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v80
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v71
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v70
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v27
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v68
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v67
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v66
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v65
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 3, v64
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v118
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v117
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v114
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v28, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v116
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v30, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v101
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 3, v99
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v28, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v29, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v112
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v103
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v97
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v102
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xff, v30
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v98
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v87
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v28, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v33, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v35, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v36, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v37, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v113, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v114, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v115, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v116, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v87, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v96, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v97, v30
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v98, v31
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v99, v32
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v27
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v2, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v1, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v30, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v34, v35
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v29, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v34, v30
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v31, v32
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-TRUE16-NEXT:  .LBB75_3: ; %end
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:320
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    s_clause 0x1f
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:320
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:352
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:444
+; GFX11-TRUE16-NEXT:    s_clause 0x7
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:448
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:452
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:456
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:460
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:464
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:468
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:472
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:476
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ; GFX11-TRUE16-NEXT:  .LBB75_4:
@@ -142181,1735 +135375,870 @@ define inreg <16 x double> @bitcast_v64bf16_to_v16f64_scalar(<64 x bfloat> inreg
 ; GFX9-NEXT:  .LBB79_4:
 ; GFX9-NEXT:    s_branch .LBB79_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v16f64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:192
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:164
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v120, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v121, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v122, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v123, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v124, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v125, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v126, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v127, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v136, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v137, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v138, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v139, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v140, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v141, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v142, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v143, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v152, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v153, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v154, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v155, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v156, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v157, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v158, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v159, s32 offset:36
-; GFX11-TRUE16-NEXT:    s_clause 0x8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v168, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v169, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v170, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v171, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v172, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v173, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v174, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v175, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v184, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v183, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v182, v9 :: v_dual_mov_b32 v169, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v171, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v176, v3 :: v_dual_mov_b32 v175, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v173, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v181, s28 :: v_dual_mov_b32 v172, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB79_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB79_3
-; GFX11-TRUE16-NEXT:  .LBB79_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s27, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s27, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s26, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s25, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s25, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff0000, v181
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v8, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v8
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v10, v6
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s24, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s23, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v7, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s22, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v11, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s22, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v9, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s21, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v14, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v11
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v16, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s20, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v12, v16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s20, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s19, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v18, v12
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v16, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s18, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v21, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v18
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s17, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v17, v20
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v19, v22
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s17, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v19, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s16, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, v24, v19
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v25, 16, 1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s16, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, v22, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v19
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, 0x400000, v25
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s3, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, v27, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v23
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v26, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v28, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s2, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, v25, v28
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v30, v25, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s1, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, v30, v25
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v31, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, v28, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v31
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v30, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v33, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, 0x400000, v29
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v30
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v29, v32, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v31, v34, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, v29, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, v31, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v31, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v37, v33
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v35, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v33, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v36, v37
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v183
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v35, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v183
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v37, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v182
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v182
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, v35, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v36, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v183, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v38, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v39, v36
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v182, v32, 16, v33
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v177
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v35, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v177
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v32, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v49, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v49, v35
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v37, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v31, v36, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v176
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v31, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v177, v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v50, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v176
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v49, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v50, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v37, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, v48, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, v39, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v176, v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v37
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v38
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v34, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v37, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v39, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, v38, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v181
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v36, v49, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v38
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, v36, v38
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, 0x400000, v48
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v50, v37, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v39, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, 0x400000, v39
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v51, v48, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, v38, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v37
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v52, v50, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, v51, v48
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v52, v52, v50
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v52, 0x400000, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v181, v39, 16, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
-; GFX11-TRUE16-NEXT:  .LBB79_3: ; %end
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v181 :: v_dual_mov_b32 v19, v175
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v21, v176
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v184, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v175, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v174, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v173, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v172, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v171, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v170, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v169, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v168, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v159, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v158, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v157, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v156, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v155, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v154, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v153, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v152, off, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v143, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v142, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v141, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v140, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v139, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v138, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v137, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v136, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v127, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v126, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v125, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v124, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v123, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v122, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v121, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v120, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:192
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    s_clause 0x8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v177 :: v_dual_mov_b32 v27, v182
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v183 :: v_dual_mov_b32 v31, v178
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v30, v179
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB79_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
-; GFX11-TRUE16-NEXT:    s_branch .LBB79_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v16f64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:288
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:284
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:280
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:276
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:272
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:268
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:264
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:260
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:256
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:252
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:248
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:244
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:240
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:236
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:232
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:228
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:224
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:220
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:216
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:212
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:208
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:204
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:200
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:196
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:192
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:188
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:184
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:180
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:176
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:172
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:168
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:164
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:160
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:156
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:152
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:148
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:144
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:140
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:136
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:132
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:128
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:124
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:120
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:116
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:112
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:108
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:104
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:100
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v136, s32 offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v137, s32 offset:92
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v138, s32 offset:88
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v139, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v140, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v141, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v142, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v143, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v152, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v153, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v154, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v155, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v156, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v157, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v158, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v159, s32 offset:36
-; GFX11-FAKE16-NEXT:    s_clause 0x8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v168, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v169, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v170, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v171, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v172, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v173, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v174, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v175, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v184, s32
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v181, v9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v182, v10 :: v_dual_mov_b32 v169, v7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v176, v6 :: v_dual_mov_b32 v171, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v173, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v175, v1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v183, s28 :: v_dual_mov_b32 v172, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB79_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB79_3
-; GFX11-FAKE16-NEXT:  .LBB79_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s27, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s27, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s26, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s25, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s25, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s24, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff0000, v183
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v8, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v3, v8
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v10, v6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s24, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s23, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v7, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s23, 16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s22, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v11, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s22, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v9, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v10, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s21, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v14, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s21, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v11
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v16, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s20, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, v12, v16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s20, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s19, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v18, v12
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s19, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v16, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s18, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s18, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v18
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s17, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v17, v20
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v19, v22
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s17, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v19, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s16, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v24, v19
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v25, 16, 1
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v22, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v19
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v23, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, 0x400000, v25
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s3, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, v27, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v23
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v24
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v26, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v25, v28, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s2, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, v25, v28
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, 0x400000, v28
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v30, v25, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s1, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, v30, v25
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v31, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, v28, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v31
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v30, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v33, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, 0x400000, v29
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, 0x400000, v30
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v29, v32, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v31, v34, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, v29, v32
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, v31, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v31, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v37, v33
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v35, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v33, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v36, v37
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v182
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v35, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v182
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v37, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v181
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v181
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v35, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v36, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v182, v31, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v38, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v39, v36
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v181, v32, 16, v33
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v176
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v35, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v176
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v32, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v49, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v37, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v31, v36, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v177
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v31, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v176, v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v177
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v50, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v37, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, v48, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, v39, v38
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v177, v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v37
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v38
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v34, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v37, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, v39, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, v38, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v36, v49 :: v_dual_lshlrev_b32 v48, 16, v183
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v38
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v36, v38
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v48
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v37, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v39, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v39
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v48, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, v38, v39
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v37
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v50, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v51, v51, v48
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v52, v52, v50
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, 0x400000, v50
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v183, v39, 16, v48
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
-; GFX11-FAKE16-NEXT:  .LBB79_3: ; %end
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v19, v175
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v184, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v175, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v174, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v173, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v172, off, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v171, off, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v170, off, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v169, off, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v168, off, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v159, off, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v158, off, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v157, off, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v156, off, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v155, off, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v154, off, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v153, off, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v152, off, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v143, off, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v142, off, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v141, off, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v140, off, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v139, off, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v138, off, s32 offset:88
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v137, off, s32 offset:92
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v136, off, s32 offset:96
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:100
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:104
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:108
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:112
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:116
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:120
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:124
-; GFX11-FAKE16-NEXT:    s_clause 0x1f
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:128
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:132
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:136
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:140
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:144
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:148
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:152
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:156
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:160
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:164
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:168
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:172
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:176
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:180
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:184
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:188
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:192
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:196
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:200
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:204
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:208
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:212
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:216
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:220
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:224
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:228
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:232
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:236
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:240
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:244
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:248
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:252
-; GFX11-FAKE16-NEXT:    s_clause 0x8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:256
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:260
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:264
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:268
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:272
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:276
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:280
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:284
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:288
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v183 :: v_dual_mov_b32 v21, v177
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, v176 :: v_dual_mov_b32 v27, v181
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v28, v182
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, v179 :: v_dual_mov_b32 v31, v178
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB79_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
-; GFX11-FAKE16-NEXT:    s_branch .LBB79_2
+; GFX11-LABEL: bitcast_v64bf16_to_v16f64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:288
+; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:284
+; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:280
+; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:276
+; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:272
+; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:268
+; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:264
+; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:260
+; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:256
+; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:252
+; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:248
+; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:244
+; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:240
+; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:236
+; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:232
+; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:228
+; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:224
+; GFX11-NEXT:    scratch_store_b32 off, v73, s32 offset:220
+; GFX11-NEXT:    scratch_store_b32 off, v74, s32 offset:216
+; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:212
+; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:208
+; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:204
+; GFX11-NEXT:    scratch_store_b32 off, v78, s32 offset:200
+; GFX11-NEXT:    scratch_store_b32 off, v79, s32 offset:196
+; GFX11-NEXT:    scratch_store_b32 off, v88, s32 offset:192
+; GFX11-NEXT:    scratch_store_b32 off, v89, s32 offset:188
+; GFX11-NEXT:    scratch_store_b32 off, v90, s32 offset:184
+; GFX11-NEXT:    scratch_store_b32 off, v91, s32 offset:180
+; GFX11-NEXT:    scratch_store_b32 off, v92, s32 offset:176
+; GFX11-NEXT:    scratch_store_b32 off, v93, s32 offset:172
+; GFX11-NEXT:    scratch_store_b32 off, v94, s32 offset:168
+; GFX11-NEXT:    scratch_store_b32 off, v95, s32 offset:164
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_store_b32 off, v104, s32 offset:160
+; GFX11-NEXT:    scratch_store_b32 off, v105, s32 offset:156
+; GFX11-NEXT:    scratch_store_b32 off, v106, s32 offset:152
+; GFX11-NEXT:    scratch_store_b32 off, v107, s32 offset:148
+; GFX11-NEXT:    scratch_store_b32 off, v108, s32 offset:144
+; GFX11-NEXT:    scratch_store_b32 off, v109, s32 offset:140
+; GFX11-NEXT:    scratch_store_b32 off, v110, s32 offset:136
+; GFX11-NEXT:    scratch_store_b32 off, v111, s32 offset:132
+; GFX11-NEXT:    scratch_store_b32 off, v120, s32 offset:128
+; GFX11-NEXT:    scratch_store_b32 off, v121, s32 offset:124
+; GFX11-NEXT:    scratch_store_b32 off, v122, s32 offset:120
+; GFX11-NEXT:    scratch_store_b32 off, v123, s32 offset:116
+; GFX11-NEXT:    scratch_store_b32 off, v124, s32 offset:112
+; GFX11-NEXT:    scratch_store_b32 off, v125, s32 offset:108
+; GFX11-NEXT:    scratch_store_b32 off, v126, s32 offset:104
+; GFX11-NEXT:    scratch_store_b32 off, v127, s32 offset:100
+; GFX11-NEXT:    scratch_store_b32 off, v136, s32 offset:96
+; GFX11-NEXT:    scratch_store_b32 off, v137, s32 offset:92
+; GFX11-NEXT:    scratch_store_b32 off, v138, s32 offset:88
+; GFX11-NEXT:    scratch_store_b32 off, v139, s32 offset:84
+; GFX11-NEXT:    scratch_store_b32 off, v140, s32 offset:80
+; GFX11-NEXT:    scratch_store_b32 off, v141, s32 offset:76
+; GFX11-NEXT:    scratch_store_b32 off, v142, s32 offset:72
+; GFX11-NEXT:    scratch_store_b32 off, v143, s32 offset:68
+; GFX11-NEXT:    scratch_store_b32 off, v152, s32 offset:64
+; GFX11-NEXT:    scratch_store_b32 off, v153, s32 offset:60
+; GFX11-NEXT:    scratch_store_b32 off, v154, s32 offset:56
+; GFX11-NEXT:    scratch_store_b32 off, v155, s32 offset:52
+; GFX11-NEXT:    scratch_store_b32 off, v156, s32 offset:48
+; GFX11-NEXT:    scratch_store_b32 off, v157, s32 offset:44
+; GFX11-NEXT:    scratch_store_b32 off, v158, s32 offset:40
+; GFX11-NEXT:    scratch_store_b32 off, v159, s32 offset:36
+; GFX11-NEXT:    s_clause 0x8
+; GFX11-NEXT:    scratch_store_b32 off, v168, s32 offset:32
+; GFX11-NEXT:    scratch_store_b32 off, v169, s32 offset:28
+; GFX11-NEXT:    scratch_store_b32 off, v170, s32 offset:24
+; GFX11-NEXT:    scratch_store_b32 off, v171, s32 offset:20
+; GFX11-NEXT:    scratch_store_b32 off, v172, s32 offset:16
+; GFX11-NEXT:    scratch_store_b32 off, v173, s32 offset:12
+; GFX11-NEXT:    scratch_store_b32 off, v174, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v175, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v184, s32
+; GFX11-NEXT:    v_dual_mov_b32 v178, v13 :: v_dual_mov_b32 v179, v12
+; GFX11-NEXT:    v_dual_mov_b32 v180, v11 :: v_dual_mov_b32 v181, v9
+; GFX11-NEXT:    v_dual_mov_b32 v182, v10 :: v_dual_mov_b32 v169, v7
+; GFX11-NEXT:    v_dual_mov_b32 v170, v8 :: v_dual_mov_b32 v177, v3
+; GFX11-NEXT:    v_dual_mov_b32 v176, v6 :: v_dual_mov_b32 v171, v4
+; GFX11-NEXT:    v_dual_mov_b32 v174, v5 :: v_dual_mov_b32 v173, v0
+; GFX11-NEXT:    v_dual_mov_b32 v184, v2 :: v_dual_mov_b32 v175, v1
+; GFX11-NEXT:    v_dual_mov_b32 v183, s28 :: v_dual_mov_b32 v172, s29
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB79_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_dual_mov_b32 v32, s0 :: v_dual_mov_b32 v37, s2
+; GFX11-NEXT:    v_dual_mov_b32 v34, s1 :: v_dual_mov_b32 v41, s3
+; GFX11-NEXT:    v_dual_mov_b32 v46, s16 :: v_dual_mov_b32 v59, s18
+; GFX11-NEXT:    v_dual_mov_b32 v52, s17 :: v_dual_mov_b32 v67, s19
+; GFX11-NEXT:    v_dual_mov_b32 v76, s20 :: v_dual_mov_b32 v97, s22
+; GFX11-NEXT:    v_dual_mov_b32 v86, s21 :: v_dual_mov_b32 v109, s23
+; GFX11-NEXT:    v_dual_mov_b32 v122, s24 :: v_dual_mov_b32 v151, s26
+; GFX11-NEXT:    v_dual_mov_b32 v136, s25 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB79_3
+; GFX11-NEXT:  .LBB79_2: ; %cmp.true
+; GFX11-NEXT:    s_and_b32 s5, s27, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s4, s27, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s6, s26, 16
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s6
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v1
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_bfe_u32 v10, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_bfe_u32 v9, v3, 16, 1
+; GFX11-NEXT:    s_lshl_b32 s7, s25, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    s_and_b32 s5, s25, 0xffff0000
+; GFX11-NEXT:    s_and_b32 s4, s24, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s5
+; GFX11-NEXT:    v_and_b32_e32 v51, 0xffff0000, v183
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v10, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v9, v3
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_bfe_u32 v10, v6, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v4, v7, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_bfe_u32 v3, v8, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v3, v8
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v10, v6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    s_lshl_b32 s4, s24, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v6
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v8
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
+; GFX11-NEXT:    s_and_b32 s4, s23, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v4, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v5
+; GFX11-NEXT:    v_dual_cndmask_b32 v6, v1, v7 :: v_dual_and_b32 v1, 0xffff, v2
+; GFX11-NEXT:    v_bfe_u32 v7, v9, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, v7, v9
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s23, 16
+; GFX11-NEXT:    v_lshl_or_b32 v151, v0, 16, v1
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_bfe_u32 v11, v7, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v5, v4, v8, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    s_and_b32 s4, s22, 0xffff0000
+; GFX11-NEXT:    v_bfe_u32 v9, v12, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v11, v7
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s22, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
+; GFX11-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v9, v12
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v7
+; GFX11-NEXT:    v_bfe_u32 v14, v10, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_or_b32_e32 v13, 0x400000, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
+; GFX11-NEXT:    s_and_b32 s4, s21, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v6, v9, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v9, v11, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, v14, v10
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v7
+; GFX11-NEXT:    v_dual_cndmask_b32 v8, v8, v13 :: v_dual_add_nc_u32 v7, v9, v11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v10
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    s_lshl_b32 s4, s21, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v11
+; GFX11-NEXT:    v_add_f32_e64 v16, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v10, v13, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_bfe_u32 v12, v16, 16, 1
+; GFX11-NEXT:    s_and_b32 s4, s20, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v11, v7, v14 :: v_dual_add_nc_u32 v10, v10, v13
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, v12, v16
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    s_lshl_b32 s4, s20, 16
+; GFX11-NEXT:    v_or_b32_e32 v17, 0x400000, v16
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
+; GFX11-NEXT:    v_bfe_u32 v18, v12, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v10, v14, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
+; GFX11-NEXT:    s_and_b32 s4, s19, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v18, v12
+; GFX11-NEXT:    v_bfe_u32 v16, v19, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v9
+; GFX11-NEXT:    v_cndmask_b32_e32 v11, v11, v17, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v17, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s19, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v16, v19
+; GFX11-NEXT:    v_or_b32_e32 v16, 0x400000, v12
+; GFX11-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v21, v17, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
+; GFX11-NEXT:    v_or_b32_e32 v20, 0x400000, v19
+; GFX11-NEXT:    s_and_b32 s4, s18, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v13, v16, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v16, v18, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
+; GFX11-NEXT:    v_add_nc_u32_e32 v19, v21, v17
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v13
+; GFX11-NEXT:    v_dual_cndmask_b32 v14, v14, v20 :: v_dual_add_nc_u32 v13, v16, v18
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
+; GFX11-NEXT:    v_or_b32_e32 v19, 0x400000, v17
+; GFX11-NEXT:    v_add_f32_e64 v20, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
+; GFX11-NEXT:    s_lshl_b32 s4, s18, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
+; GFX11-NEXT:    v_or_b32_e32 v21, 0x400000, v18
+; GFX11-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v17, v20, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_bfe_u32 v19, v22, 16, 1
+; GFX11-NEXT:    s_and_b32 s4, s17, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v17, v17, v20
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v18, v13, v21 :: v_dual_and_b32 v13, 0xffff, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v16
+; GFX11-NEXT:    v_or_b32_e32 v21, 0x400000, v20
+; GFX11-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
+; GFX11-NEXT:    v_add_nc_u32_e32 v18, v19, v22
+; GFX11-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
+; GFX11-NEXT:    s_lshl_b32 s4, s17, 16
+; GFX11-NEXT:    v_or_b32_e32 v23, 0x400000, v22
+; GFX11-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
+; GFX11-NEXT:    v_bfe_u32 v24, v19, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v25, 0x40c00000, s4
+; GFX11-NEXT:    v_cndmask_b32_e32 v20, v17, v21, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
+; GFX11-NEXT:    s_and_b32 s4, s16, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, v24, v19
+; GFX11-NEXT:    v_bfe_u32 v22, v25, 16, 1
+; GFX11-NEXT:    v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v17, 0xffff, v16
+; GFX11-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s16, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
+; GFX11-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v21
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, v22, v25
+; GFX11-NEXT:    v_or_b32_e32 v22, 0x400000, v19
+; GFX11-NEXT:    v_add_f32_e64 v24, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v27, v23, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
+; GFX11-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
+; GFX11-NEXT:    v_or_b32_e32 v26, 0x400000, v25
+; GFX11-NEXT:    s_and_b32 s4, s3, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v20, v20, v22, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v22, v24, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
+; GFX11-NEXT:    v_add_nc_u32_e32 v25, v27, v23
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v20
+; GFX11-NEXT:    v_dual_cndmask_b32 v21, v21, v26 :: v_dual_add_nc_u32 v20, v22, v24
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v25
+; GFX11-NEXT:    v_or_b32_e32 v25, 0x400000, v23
+; GFX11-NEXT:    v_add_f32_e64 v26, 0x40c00000, s4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
+; GFX11-NEXT:    v_or_b32_e32 v27, 0x400000, v24
+; GFX11-NEXT:    v_add_f32_e64 v28, 0x40c00000, s3
+; GFX11-NEXT:    v_cndmask_b32_e32 v22, v22, v25, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v23, v26, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
+; GFX11-NEXT:    v_bfe_u32 v25, v28, 16, 1
+; GFX11-NEXT:    s_and_b32 s3, s2, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v24, v20, v27 :: v_dual_add_nc_u32 v23, v23, v26
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v22
+; GFX11-NEXT:    v_or_b32_e32 v27, 0x400000, v26
+; GFX11-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v24, v25, v28
+; GFX11-NEXT:    v_add_f32_e64 v25, 0x40c00000, s3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_or_b32_e32 v29, 0x400000, v28
+; GFX11-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
+; GFX11-NEXT:    v_bfe_u32 v30, v25, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v31, 0x40c00000, s2
+; GFX11-NEXT:    v_cndmask_b32_e32 v26, v23, v27, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
+; GFX11-NEXT:    s_and_b32 s2, s1, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, v30, v25
+; GFX11-NEXT:    v_bfe_u32 v28, v31, 16, 1
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v24, v24, v29, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v29, 0x40c00000, s2
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v26
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v27
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, v28, v31
+; GFX11-NEXT:    v_or_b32_e32 v28, 0x400000, v25
+; GFX11-NEXT:    v_add_f32_e64 v30, 0x40c00000, s1
+; GFX11-NEXT:    v_bfe_u32 v33, v29, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
+; GFX11-NEXT:    v_or_b32_e32 v32, 0x400000, v31
+; GFX11-NEXT:    s_and_b32 s1, s0, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v26, v26, v28, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v28, v30, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v33, v29
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v26
+; GFX11-NEXT:    v_dual_cndmask_b32 v27, v27, v32 :: v_dual_add_nc_u32 v26, v28, v30
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v31
+; GFX11-NEXT:    v_or_b32_e32 v31, 0x400000, v29
+; GFX11-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, 0x7fff, v26
+; GFX11-NEXT:    v_or_b32_e32 v33, 0x400000, v30
+; GFX11-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
+; GFX11-NEXT:    v_cndmask_b32_e32 v28, v28, v31, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v29, v32, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
+; GFX11-NEXT:    v_or_b32_e32 v35, 0x400000, v32
+; GFX11-NEXT:    v_bfe_u32 v31, v34, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
+; GFX11-NEXT:    v_cndmask_b32_e32 v30, v26, v33, vcc_lo
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v28
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, v29, v32
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v178
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v30
+; GFX11-NEXT:    v_add_nc_u32_e32 v30, v31, v34
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff0000, v178
+; GFX11-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
+; GFX11-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
+; GFX11-NEXT:    v_lshl_or_b32 v109, v5, 16, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v30, 0x7fff, v30
+; GFX11-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v28, v35, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v37, v33, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v29
+; GFX11-NEXT:    v_bfe_u32 v35, v31, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v37, v33
+; GFX11-NEXT:    v_cndmask_b32_e32 v30, v30, v36, vcc_lo
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v179
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v35, v31
+; GFX11-NEXT:    v_or_b32_e32 v37, 0x400000, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v179
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_or_b32_e32 v38, 0x400000, v31
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v32, v37, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v180
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v34, v38, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v34, v36, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v33, v35, 16, 1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v180
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v33, v35
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_bfe_u32 v36, v37, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_bfe_u32 v35, v38, 16, 1
+; GFX11-NEXT:    v_lshl_or_b32 v178, v31, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v36, v37
+; GFX11-NEXT:    v_dual_cndmask_b32 v33, v33, v48 :: v_dual_lshlrev_b32 v36, 16, v182
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v35, v38
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v182
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_lshl_or_b32 v179, v32, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xffff, v30
+; GFX11-NEXT:    v_lshl_or_b32 v136, v2, 16, v4
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_bfe_u32 v37, v36, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_dual_cndmask_b32 v31, v31, v48 :: v_dual_add_nc_u32 v38, v38, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v37, v36
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v181
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v181
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v32, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-NEXT:    v_bfe_u32 v35, v37, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v36, v38, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v180, v31, 16, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, v35, v37
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v170
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v36, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v36, 16, v170
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v37
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v39, v36, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_lshl_or_b32 v182, v31, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v38, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v39, v36
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v48, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v38, 0x400000, v36
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v37
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v169
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v31, v38, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v169
+; GFX11-NEXT:    v_lshl_or_b32 v181, v32, 16, v33
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff0000, v176
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
+; GFX11-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_bfe_u32 v35, v37, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v34
+; GFX11-NEXT:    v_bfe_u32 v32, v36, 16, 1
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v35, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 16, v176
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v32, v36
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
+; GFX11-NEXT:    v_bfe_u32 v37, v38, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v34, v39, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v49, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_lshl_or_b32 v170, v33, 16, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v49, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v32, v32, v48 :: v_dual_add_nc_u32 v33, v37, v38
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v174
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v36
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v34, v34, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v174
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v37, v35, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_cndmask_b32 v33, v33, v39
+; GFX11-NEXT:    v_lshl_or_b32 v169, v31, 16, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v37, v35
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v31, v36, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff0000, v171
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 16, v177
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v31, v36
+; GFX11-NEXT:    v_lshl_or_b32 v176, v33, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v37
+; GFX11-NEXT:    v_or_b32_e32 v34, 0x400000, v35
+; GFX11-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v37, 16, v171
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v34, 0x400000, v36
+; GFX11-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
+; GFX11-NEXT:    v_bfe_u32 v37, v32, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v32
+; GFX11-NEXT:    v_bfe_u32 v50, v38, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v48, 0x400000, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v31, v31, v34 :: v_dual_add_nc_u32 v36, v37, v32
+; GFX11-NEXT:    v_bfe_u32 v34, v35, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v177
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_nc_u32 v34, v34, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v32, v36, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_bfe_u32 v49, v37, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v50, v38
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v50, 16, v184
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v34, v34, v48 :: v_dual_add_nc_u32 v35, v49, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff0000, v184
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v37
+; GFX11-NEXT:    v_cndmask_b32_e32 v36, v36, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v50
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_dual_add_f32 v38, 0x40c00000, v48 :: v_dual_cndmask_b32 v35, v35, v49
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_bfe_u32 v48, v37, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v39, v38, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-NEXT:    v_lshl_or_b32 v174, v33, 16, v31
+; GFX11-NEXT:    v_lshl_or_b32 v171, v32, 16, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, v48, v37
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff0000, v175
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 16, v175
+; GFX11-NEXT:    v_add_nc_u32_e32 v39, v39, v38
+; GFX11-NEXT:    v_lshl_or_b32 v177, v35, 16, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v31, 0x7fff, v31
+; GFX11-NEXT:    v_or_b32_e32 v35, 0x400000, v37
+; GFX11-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v39
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v38
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v37, v33, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v39, v34, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v31, v31, v35, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v35, 0xffff0000, v173
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v173
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
+; GFX11-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v37, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, v39, v34
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v38, v35, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, v38, v35
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_lshl_or_b32 v122, v3, 16, v6
+; GFX11-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v38
+; GFX11-NEXT:    v_add_f32_e32 v38, 0x40c00000, v48
+; GFX11-NEXT:    v_or_b32_e32 v39, 0x400000, v35
+; GFX11-NEXT:    v_dual_cndmask_b32 v33, v36, v49 :: v_dual_lshlrev_b32 v48, 16, v183
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v36, v38, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v49, 0x400000, v38
+; GFX11-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
+; GFX11-NEXT:    v_cndmask_b32_e32 v35, v37, v39, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff0000, v172
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v172
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, v36, v38
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
+; GFX11-NEXT:    v_or_b32_e32 v55, 0x400000, v48
+; GFX11-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
+; GFX11-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    v_bfe_u32 v50, v37, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v38, v39, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v36, v36, v49, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v54, 0x400000, v39
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
+; GFX11-NEXT:    v_dual_add_f32 v50, 0x40c00000, v51 :: v_dual_add_nc_u32 v49, v50, v37
+; GFX11-NEXT:    v_bfe_u32 v51, v48, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, v38, v39
+; GFX11-NEXT:    v_or_b32_e32 v53, 0x400000, v37
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v49, 0x7fff, v49
+; GFX11-NEXT:    v_bfe_u32 v52, v50, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v51, v51, v48
+; GFX11-NEXT:    v_add_nc_u32_e32 v38, 0x7fff, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v52, v52, v50
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v38, v38, v54 :: v_dual_add_nc_u32 v51, 0x7fff, v51
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v52
+; GFX11-NEXT:    v_or_b32_e32 v52, 0x400000, v50
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
+; GFX11-NEXT:    v_cndmask_b32_e32 v48, v51, v55, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-NEXT:    v_lshl_or_b32 v184, v32, 16, v31
+; GFX11-NEXT:    v_lshl_or_b32 v175, v33, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v48
+; GFX11-NEXT:    v_cndmask_b32_e32 v37, v49, v53, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
+; GFX11-NEXT:    v_lshl_or_b32 v173, v35, 16, v36
+; GFX11-NEXT:    v_lshl_or_b32 v97, v8, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v48
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v37
+; GFX11-NEXT:    v_cndmask_b32_e32 v39, v39, v52, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v86, v9, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v76, v11, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v67, v14, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v172, v37, 16, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v39
+; GFX11-NEXT:    v_lshl_or_b32 v59, v16, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v52, v18, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v46, v21, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v41, v22, 16, v25
+; GFX11-NEXT:    v_lshl_or_b32 v183, v39, 16, v48
+; GFX11-NEXT:    v_lshl_or_b32 v37, v24, 16, v27
+; GFX11-NEXT:    v_lshl_or_b32 v34, v26, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v32, v29, 16, v30
+; GFX11-NEXT:  .LBB79_3: ; %end
+; GFX11-NEXT:    v_dual_mov_b32 v3, v41 :: v_dual_mov_b32 v4, v46
+; GFX11-NEXT:    v_dual_mov_b32 v6, v59 :: v_dual_mov_b32 v9, v86
+; GFX11-NEXT:    v_dual_mov_b32 v7, v67 :: v_dual_mov_b32 v8, v76
+; GFX11-NEXT:    v_dual_mov_b32 v10, v97 :: v_dual_mov_b32 v13, v136
+; GFX11-NEXT:    v_dual_mov_b32 v11, v109 :: v_dual_mov_b32 v12, v122
+; GFX11-NEXT:    v_dual_mov_b32 v14, v151 :: v_dual_mov_b32 v17, v172
+; GFX11-NEXT:    v_dual_mov_b32 v18, v173 :: v_dual_mov_b32 v19, v175
+; GFX11-NEXT:    v_dual_mov_b32 v20, v184 :: v_dual_mov_b32 v23, v174
+; GFX11-NEXT:    v_dual_mov_b32 v22, v171 :: v_dual_mov_b32 v25, v169
+; GFX11-NEXT:    v_dual_mov_b32 v26, v170 :: v_dual_mov_b32 v29, v180
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_load_b32 v184, off, s32
+; GFX11-NEXT:    scratch_load_b32 v175, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v174, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v173, off, s32 offset:12
+; GFX11-NEXT:    scratch_load_b32 v172, off, s32 offset:16
+; GFX11-NEXT:    scratch_load_b32 v171, off, s32 offset:20
+; GFX11-NEXT:    scratch_load_b32 v170, off, s32 offset:24
+; GFX11-NEXT:    scratch_load_b32 v169, off, s32 offset:28
+; GFX11-NEXT:    scratch_load_b32 v168, off, s32 offset:32
+; GFX11-NEXT:    scratch_load_b32 v159, off, s32 offset:36
+; GFX11-NEXT:    scratch_load_b32 v158, off, s32 offset:40
+; GFX11-NEXT:    scratch_load_b32 v157, off, s32 offset:44
+; GFX11-NEXT:    scratch_load_b32 v156, off, s32 offset:48
+; GFX11-NEXT:    scratch_load_b32 v155, off, s32 offset:52
+; GFX11-NEXT:    scratch_load_b32 v154, off, s32 offset:56
+; GFX11-NEXT:    scratch_load_b32 v153, off, s32 offset:60
+; GFX11-NEXT:    scratch_load_b32 v152, off, s32 offset:64
+; GFX11-NEXT:    scratch_load_b32 v143, off, s32 offset:68
+; GFX11-NEXT:    scratch_load_b32 v142, off, s32 offset:72
+; GFX11-NEXT:    scratch_load_b32 v141, off, s32 offset:76
+; GFX11-NEXT:    scratch_load_b32 v140, off, s32 offset:80
+; GFX11-NEXT:    scratch_load_b32 v139, off, s32 offset:84
+; GFX11-NEXT:    scratch_load_b32 v138, off, s32 offset:88
+; GFX11-NEXT:    scratch_load_b32 v137, off, s32 offset:92
+; GFX11-NEXT:    scratch_load_b32 v136, off, s32 offset:96
+; GFX11-NEXT:    scratch_load_b32 v127, off, s32 offset:100
+; GFX11-NEXT:    scratch_load_b32 v126, off, s32 offset:104
+; GFX11-NEXT:    scratch_load_b32 v125, off, s32 offset:108
+; GFX11-NEXT:    scratch_load_b32 v124, off, s32 offset:112
+; GFX11-NEXT:    scratch_load_b32 v123, off, s32 offset:116
+; GFX11-NEXT:    scratch_load_b32 v122, off, s32 offset:120
+; GFX11-NEXT:    scratch_load_b32 v121, off, s32 offset:124
+; GFX11-NEXT:    s_clause 0x1f
+; GFX11-NEXT:    scratch_load_b32 v120, off, s32 offset:128
+; GFX11-NEXT:    scratch_load_b32 v111, off, s32 offset:132
+; GFX11-NEXT:    scratch_load_b32 v110, off, s32 offset:136
+; GFX11-NEXT:    scratch_load_b32 v109, off, s32 offset:140
+; GFX11-NEXT:    scratch_load_b32 v108, off, s32 offset:144
+; GFX11-NEXT:    scratch_load_b32 v107, off, s32 offset:148
+; GFX11-NEXT:    scratch_load_b32 v106, off, s32 offset:152
+; GFX11-NEXT:    scratch_load_b32 v105, off, s32 offset:156
+; GFX11-NEXT:    scratch_load_b32 v104, off, s32 offset:160
+; GFX11-NEXT:    scratch_load_b32 v95, off, s32 offset:164
+; GFX11-NEXT:    scratch_load_b32 v94, off, s32 offset:168
+; GFX11-NEXT:    scratch_load_b32 v93, off, s32 offset:172
+; GFX11-NEXT:    scratch_load_b32 v92, off, s32 offset:176
+; GFX11-NEXT:    scratch_load_b32 v91, off, s32 offset:180
+; GFX11-NEXT:    scratch_load_b32 v90, off, s32 offset:184
+; GFX11-NEXT:    scratch_load_b32 v89, off, s32 offset:188
+; GFX11-NEXT:    scratch_load_b32 v88, off, s32 offset:192
+; GFX11-NEXT:    scratch_load_b32 v79, off, s32 offset:196
+; GFX11-NEXT:    scratch_load_b32 v78, off, s32 offset:200
+; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:204
+; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:208
+; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:212
+; GFX11-NEXT:    scratch_load_b32 v74, off, s32 offset:216
+; GFX11-NEXT:    scratch_load_b32 v73, off, s32 offset:220
+; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:224
+; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:228
+; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:232
+; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:236
+; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:240
+; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:244
+; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:248
+; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:252
+; GFX11-NEXT:    s_clause 0x8
+; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:256
+; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:260
+; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:264
+; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:268
+; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:272
+; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:276
+; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:280
+; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:284
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:288
+; GFX11-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v34
+; GFX11-NEXT:    v_dual_mov_b32 v2, v37 :: v_dual_mov_b32 v5, v52
+; GFX11-NEXT:    v_dual_mov_b32 v16, v183 :: v_dual_mov_b32 v21, v177
+; GFX11-NEXT:    v_dual_mov_b32 v24, v176 :: v_dual_mov_b32 v27, v181
+; GFX11-NEXT:    v_mov_b32_e32 v28, v182
+; GFX11-NEXT:    v_dual_mov_b32 v30, v179 :: v_dual_mov_b32 v31, v178
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB79_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
+; GFX11-NEXT:    ; implicit-def: $vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73
+; GFX11-NEXT:    ; implicit-def: $vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78
+; GFX11-NEXT:    ; implicit-def: $vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84
+; GFX11-NEXT:    ; implicit-def: $vgpr60_vgpr61_vgpr62_vgpr63_vgpr64_vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91
+; GFX11-NEXT:    ; implicit-def: $vgpr68_vgpr69_vgpr70_vgpr71_vgpr72_vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99
+; GFX11-NEXT:    ; implicit-def: $vgpr77_vgpr78_vgpr79_vgpr80_vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108
+; GFX11-NEXT:    ; implicit-def: $vgpr87_vgpr88_vgpr89_vgpr90_vgpr91_vgpr92_vgpr93_vgpr94_vgpr95_vgpr96_vgpr97_vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118
+; GFX11-NEXT:    ; implicit-def: $vgpr98_vgpr99_vgpr100_vgpr101_vgpr102_vgpr103_vgpr104_vgpr105_vgpr106_vgpr107_vgpr108_vgpr109_vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129
+; GFX11-NEXT:    ; implicit-def: $vgpr110_vgpr111_vgpr112_vgpr113_vgpr114_vgpr115_vgpr116_vgpr117_vgpr118_vgpr119_vgpr120_vgpr121_vgpr122_vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141
+; GFX11-NEXT:    ; implicit-def: $vgpr123_vgpr124_vgpr125_vgpr126_vgpr127_vgpr128_vgpr129_vgpr130_vgpr131_vgpr132_vgpr133_vgpr134_vgpr135_vgpr136_vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154
+; GFX11-NEXT:    ; implicit-def: $vgpr137_vgpr138_vgpr139_vgpr140_vgpr141_vgpr142_vgpr143_vgpr144_vgpr145_vgpr146_vgpr147_vgpr148_vgpr149_vgpr150_vgpr151_vgpr152_vgpr153_vgpr154_vgpr155_vgpr156_vgpr157_vgpr158_vgpr159_vgpr160_vgpr161_vgpr162_vgpr163_vgpr164_vgpr165_vgpr166_vgpr167_vgpr168
+; GFX11-NEXT:    s_branch .LBB79_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -160846,417 +153175,438 @@ define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a
 ; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64bf16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:436
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:320
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v6 :: v_dual_mov_b32 v33, v0
+; GFX11-TRUE16-NEXT:    s_clause 0x1e
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:352
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:320
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, v30 :: v_dual_mov_b32 v51, v24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v28 :: v_dual_mov_b32 v55, v26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v22 :: v_dual_mov_b32 v48, v20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v18 :: v_dual_mov_b32 v39, v16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v14 :: v_dual_mov_b32 v34, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v12 :: v_dual_mov_b32 v38, v10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v6 :: v_dual_mov_b32 v33, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v2 :: v_dual_mov_b32 v32, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
 ; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:316
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v89, off, s32 offset:312
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v92, off, s32 offset:308
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v91, off, s32 offset:304
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v93, off, s32 offset:300
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:296
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:292
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v88, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v90, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v2, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v16, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v18, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v20, off, s32 offset:72
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v22, off, s32 offset:80
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v24, off, s32 offset:88
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v26, off, s32 offset:96
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v28, off, s32 offset:104
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v30, off, s32 offset:112
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:120
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:128
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:136
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:144
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:152
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:160
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:168
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:176
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:184
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:200
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:208
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:216
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:224
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:232
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:240
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:248
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:256
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v88, off, s32 offset:264
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v89, off, s32 offset:272
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v90, off, s32 offset:280
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v91, off, s32 offset:288
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v92, off, s32 offset:296
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v93, off, s32 offset:304
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v94, off, s32 offset:312
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:308
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:300
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:292
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:284
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:276
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:268
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:260
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:252
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:244
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:228
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:220
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:212
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:204
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:196
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:188
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:180
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:172
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:164
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:156
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:148
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:140
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:132
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:124
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:116
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:108
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:100
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:92
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:84
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:76
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:68
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:12
 ; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v86, v30 :: v_dual_mov_b32 v81, v29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v84, v28 :: v_dual_mov_b32 v83, v27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v85, v26 :: v_dual_mov_b32 v80, v24
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v69, v25 :: v_dual_mov_b32 v82, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v71, v23 :: v_dual_mov_b32 v64, v21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v67, v20 :: v_dual_mov_b32 v68, v19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v70, v18 :: v_dual_mov_b32 v55, v16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v17 :: v_dual_mov_b32 v65, v15
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v66, v14 :: v_dual_mov_b32 v49, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v12 :: v_dual_mov_b32 v54, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v11 :: v_dual_mov_b32 v36, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v8 :: v_dual_mov_b32 v37, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v7 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v32, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v101, 8, v6
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v102, 8, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v113, 8, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v112, 8, v12
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v130, 8, v14
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v128, 8, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v134, 8, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v132, 8, v20
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v161, 8, v22
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v147, 8, v24
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v166, 8, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v167, 8, v28
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v180, 8, v30
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v177, 8, v31
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v42, 8, v41
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(61)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 8, v44
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(60)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v45, 8, v45
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(59)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v44, 8, v56
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(58)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v59, 8, v59
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(57)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v56, 8, v60
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(56)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v60, 8, v61
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(55)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v61, 8, v62
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(54)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v63, 8, v63
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(53)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v62, 8, v72
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(52)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v73, 8, v73
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(51)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v72, 8, v74
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(50)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v75, 8, v75
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(49)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v74, 8, v76
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(48)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v77, 8, v77
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(47)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v76, 8, v78
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(46)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v78, 8, v79
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(45)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v79, 8, v88
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(44)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v89, 8, v89
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(43)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v88, 8, v90
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(42)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v91, 8, v91
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(41)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v90, 8, v92
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(40)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v92, 8, v93
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(39)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v93, 8, v94
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB89_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v33
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v32
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v35
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v2, 0xffff, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v1, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v53
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v6, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v49
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v66
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v65
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v55
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v68
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v3, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v2, 16, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v83
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v86
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v98
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v96
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v102
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v101
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v99
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v114
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v112
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v103
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v118
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v115
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v116
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v113
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v130
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v119
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v14, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v134
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v133
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v128
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v117
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v146
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v145
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v132
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v135
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v131
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v129
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v144
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v150
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v148
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v21
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v147
+; GFX11-TRUE16-NEXT:    v_and_b32_e64 v5, 0xffff, s5
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
 ; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v149
 ; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v151
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v160
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v161
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v163
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s26, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v164
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s27, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
 ; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v165
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v178
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v177
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v68
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v64
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v4, v67
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v37
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v70
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v7, v80
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v8, v81
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v9, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v86
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v83
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v96
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v85
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v10, v97
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v11, v87
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v99
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v3, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v103
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v114
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v98
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v0, 16, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v100
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v113
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v116
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v14, v128
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v117
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v102
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v130
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v133
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v14, v132
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v0, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v148
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v118
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v129
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v161
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v2, 16, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v166
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v144
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v134
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v147
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v167
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v17, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v22
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v151
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v149
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v20, 16, v21
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v167
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v177
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v165
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v162
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v42
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v41
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v178
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v115
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v45
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v44
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v131
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v119
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v59
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v56
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v176
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v182
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v181
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v179
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v135
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v60
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v61
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v180
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v42
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v40
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v183
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v150
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v146
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v63
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v62
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v41
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v43
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v44
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v163
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v160
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v73
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v72
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v45
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v47
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v56
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v176
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v164
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v75
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v74
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v57
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v62
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v61
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v59
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v179
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v77
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v76
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v60
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v74
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v73
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v63
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v183
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v182
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v78
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v79
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v72
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v77
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v76
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v75
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v43
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v40
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v89
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v88
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v78
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v90
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v88
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v79
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v47
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v46
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v91
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v90
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v89
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v93
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v91
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v92
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v58
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v57
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v92
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v93
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v1, 16, v0
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
@@ -161303,396 +153653,344 @@ define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v92
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v93
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(37)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v58
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v57
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(35)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v47
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v90
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v79
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v92, v0
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v91
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v88
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v78
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v77
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v7, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v46
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v93, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v91, v2
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(33)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v43
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v5, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v76
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v75
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v74
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v72
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v73
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v40
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v90, v3
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(31)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v183
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v182
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v63
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(59)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v62
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v61
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v59
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v59, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v60, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v57
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(55)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v58
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v57, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v56
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v56, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v47
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v89, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(29)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v181
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v88, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v181, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v78, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v79, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v179
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v182, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v179, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v77, v3
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(27)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v176
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v45
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(51)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v46
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v45, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v44
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v43
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v41
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(47)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v42
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v164
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(25)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v163
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v163, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v76, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v160
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v75, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v160, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v74, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v73, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v1
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(23)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v150
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v40
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v183
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(43)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v182
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v180
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v181
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(39)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v178
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v176
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v72, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v146
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v177
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v167
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v167, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v176, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v165
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(35)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v166
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v165, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v164
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v164, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v163
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(21)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v145
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v135
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v63, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(19)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v131
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v62, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v131, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v60, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v61, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v119
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v135, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v119, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v59, v3
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(17)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v178
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v161
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v162
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v161, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v160
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v151
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v149
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v150
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v115
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v165
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v115, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v56, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v162
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v45, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v145, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v44, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v42, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v1
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v151
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v148
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v147
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(23)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v146
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v144
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v145
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v135
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(19)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v134
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v132
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v41, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v149
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v133
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v131
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v131, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v132, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v129
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v130
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v129, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v128
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v128, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v119
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v117
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(11)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v118
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v117, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v116
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v115
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v113
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v148
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v144
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v180, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(9)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v133
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v177, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v166, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v167, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v133, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v129
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v129, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v144, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v161, v3
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v114
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v112
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v103
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v102
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v101
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v99
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v86
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v118
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v117
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(5)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v116
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v116, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v147, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v114
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v96
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v55, 3, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v99
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v114, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v132, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v130, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v103
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v98
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v54
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v39
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 3, v33
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v113, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v128, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v100
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v101, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v102, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v98
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v96
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v134, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v97, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v87
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v85
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v84
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v83
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v81
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v82
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v87, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v51
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v86, v4
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v69
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v70
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v70, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v69, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v68
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v67
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v66
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v64
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v85, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v84, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v49
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v83, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v48
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v65
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v64, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v52
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v82, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v38
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v81, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v71, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v80, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v37
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v70, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v36
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, 3, v51
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v53, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v52, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v49
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v50
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 3, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v38, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v36, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v37, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v33
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v35
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v69, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v112, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v68, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v67, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v66, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v33
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v34, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v32
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v65, v8
 ; GFX11-TRUE16-NEXT:    v_and_b32_e64 v8, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v64
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v64, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v7, 16, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v32, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v51
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v38
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v53, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v49, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v69
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v33, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v34, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v50, 16, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v132
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v3, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v116
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v129
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v18
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v114, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v144, 16, v33
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v128
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v129, 16, v33
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v176
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v117, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v164
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v165, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v161, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v56
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v14, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v115
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v135
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v131
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v23
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v27
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v145, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v119, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v163
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v182
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v28
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v2, 16, v0
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v131, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v167, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v45, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v57, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v59, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v133, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v160, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v179, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v36
 ; GFX11-TRUE16-NEXT:  .LBB89_3: ; %end
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:320
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    s_clause 0x1e
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:320
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:352
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:440
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ; GFX11-TRUE16-NEXT:  .LBB89_4:
@@ -175425,3233 +167723,1575 @@ define inreg <128 x i8> @bitcast_v64bf16_to_v128i8_scalar(<64 x bfloat> inreg %a
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v128i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:12
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v41, s96, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s72, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s73, v2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v41, s97, 1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v41, s98, 2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v41, s99, 3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s44, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s45, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v41, s100, 4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s43, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v41, s101, 5
-; GFX11-TRUE16-NEXT:    s_mov_b32 vcc_hi, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr43 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v41, s102, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v41, s103, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v41, s104, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s50, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s51, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s52, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s53, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s54, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s55, 15
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s64, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s65, 17
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s66, 18
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s67, 19
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s68, 20
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s69, 21
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s70, 22
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s71, 23
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s80, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s81, 25
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s82, 26
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s83, 27
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s84, 28
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s85, 29
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s86, 30
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s87, 31
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB91_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[24:25], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s1, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s0, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s41, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s87, s41, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s55, s41, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s69, s40, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s70, s40, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s43, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s96, s43, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s43, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s71, s42, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s45, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s45, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s64, s45, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s80, s44, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s81, s44, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s51, s59, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 31
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s24, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s14, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s59, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s65, s59, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 30
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s23, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s15, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s82, s58, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 29
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s58, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s61, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s61, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 28
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s66, s61, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s83, s60, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s84, s60, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 27
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s52, s63, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s63, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s67, s63, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s85, s62, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s53, s73, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s97, s73, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 23
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s73, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s72, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s72, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 22
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s54, s29, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s98, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s68, s29, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 21
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s86, s28, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[100:101], s[26:27], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 20
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[46:47], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[56:57], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 17
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[88:89], s[58:59], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[90:91], s[60:61], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[62:63], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[72:73], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 15
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 14
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s19, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 13
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s18, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s18, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s17, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s17, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s17, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s62, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s14, 25
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s15, 26
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s14, 18
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s15, 19
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[18:19], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s14, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s15, 12
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB91_4
-; GFX11-TRUE16-NEXT:  .LBB91_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s29, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, s58, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s1, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s15, s58, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 16
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s41, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s41, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s77, s28, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s6, s58, 0x10010
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s78, s28, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s6, s58
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s73, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s76, s73, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s74, s72, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s75, s72, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, s63, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s73, s63, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s63, s62, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s72, s62, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s62, s61, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s57, s61, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s47, s60, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s56, s60, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s46, s59, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s28, s59, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, s45, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s45, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s45, s44, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s44, s44, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s29, s43, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s43, s43, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s13, s42, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s42, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s40, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s40, 16
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s58, 22
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s40, s58, s41
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s40, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s77
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s1, v3
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s78
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v4, v5, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s1, 0x10010
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s1
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s1, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s1, s1, s41
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v2
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s40
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v4, v6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v2
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v51, 0x40c00000, s6
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s40, 0x10010
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s40
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s40, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v6
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s40, s40, s41
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s76
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v3, v5, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v6, 16, 1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s5, s0, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v4
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s5, s0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s40, 16
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s0, 22
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, s3, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s40
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v9
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v1, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v7, 16, v3
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s40, 0x10010
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v6
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s40
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s40, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s40, s40, s41
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s75
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s40, 16
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s3, v10
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s74
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s3, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s3
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s3, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v8, v9
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s3, s3, s41
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, s2, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s40
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s40, 0x10010
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s40
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s40, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s40, s40, s41
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s73
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s2, v3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v4, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v8, 16, 1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s12, s2, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v4
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s12, s2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s40, 16
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s2, 22
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s2, s2, s41
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, s17, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v24, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s40
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v9, 16, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v5, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v4
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s40, 0x10010
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s40
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s40, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v11, v5
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s40, s40, s41
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s17, s17, 16
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s72
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s17, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s63
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v10, v11, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s17, 0x10010
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s17
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s40, 16
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s17, 22
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s17, s17, s41
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, s16, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v9, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v10, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s40
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v25, 16, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s17, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s40, 0x10010
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s40
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s40, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s62
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s40, s40, s41
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s16, s16, 16
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s16
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s40, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v11, vcc_lo
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s16, v10
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v9, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s57
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s16, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s16
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s16, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s16, s16, s41
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, s19, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s40
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v8, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s16, 16
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v12
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v6, 16, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s47
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s40, 0x10010
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v11, v8
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s40
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s40, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s40, s40, s41
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s19, s19, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s56
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s19, v12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s40, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s19, 0x10010
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v11, v10
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s19, 22
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s19, s19, s41
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, s18, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v7, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s40
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s19, s19, 16
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v8
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s46
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s15
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s41, s40, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v11, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, s40
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s40, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s41, 0x7fff
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s28
-; GFX11-TRUE16-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s40, s40, s41
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s18, s18, 16
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s18
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s40, 16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s18, v9
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v28, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v5, 16, v10
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s28, s18, 0x10010
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v7
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, s18
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s18, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s28, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s18, s18, s28
-; GFX11-TRUE16-NEXT:    s_and_b32 s28, s21, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s28
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s14
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v11, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s18, s18, 16
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v11
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v11
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s15, s14, 0x10010
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v10, 16, 1
-; GFX11-TRUE16-NEXT:    s_add_i32 s15, s15, s14
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s14, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s15, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s28, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s14, s14, s15
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s15, s21, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s15
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s14, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v13
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v8, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v9, v12
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s21, s15, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, s15
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s15, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s21, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s14, s15, s21
-; GFX11-TRUE16-NEXT:    s_and_b32 s15, s20, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s15
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s14, 16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s0, s5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s11
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s40, s2, s12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s15, s11, 0x10010
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v33
-; GFX11-TRUE16-NEXT:    s_add_i32 s15, s15, s11
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s11, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s15, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s11, s11, s15
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s14, s20, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s14
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v6, 16, v11
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s9, s14, 0x10010
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    s_add_i32 s15, s9, s14
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s11, 16
-; GFX11-TRUE16-NEXT:    s_addk_i32 s15, 0x7fff
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s14, 22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v7
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s11, s14, s15
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, s23, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s11, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s14
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s45
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s44
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s15, s14, 0x10010
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v10, 16, 1
-; GFX11-TRUE16-NEXT:    s_add_i32 s15, s15, s14
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s14, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s15, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s11, s14, s15
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v8, v10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s11, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v9, v11
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s15, s14, 0x10010
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v11
-; GFX11-TRUE16-NEXT:    s_add_i32 s15, s15, s14
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s14, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s15, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s11, s14, s15
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, s22, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s11, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s43
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s14
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s29
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v7
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v71, v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s13
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s15, s14, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v5
-; GFX11-TRUE16-NEXT:    s_add_i32 s15, s15, s14
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s14, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s15, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s11, s14, s15
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s14, s22, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s14
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v11, v10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s11, 16
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v12
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v70, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s13, s14, 0x10010
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v10
-; GFX11-TRUE16-NEXT:    s_add_i32 s13, s13, s14
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s14, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s13, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s15, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s13, s14, s13
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, s25, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s10
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v12
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s13, 16
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v6
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s14, s10, 0x10010
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v11, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, s10
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s10, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s14, 0x7fff
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
-; GFX11-TRUE16-NEXT:    s_and_b32 s13, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s10, s10, s14
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s13, s25, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v7, v8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s10, 16
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v10
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s13, s8, 0x10010
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v11, 16, 1
-; GFX11-TRUE16-NEXT:    s_add_i32 s13, s13, s8
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s13, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s8, s13
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s24, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s8, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v6, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s10
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v48, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v13, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s10, s7, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
-; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, s7
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s7, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s7, s7, s10
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s24, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s8
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v10, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s7, 16
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v12
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v7, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v10
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v51, 16, 1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s4, s8, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v6
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, s8
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s4, s8, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s27, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v52, 0x40c00000, s6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v12, v51
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s4, 16
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v11
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v11, 16, 1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s7, s6, 0x10010
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v50
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, s6
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s6, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s4, s6, s7
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v8, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v51
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v51, v51
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v66, v5, 16, v13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v10
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v67, v38, 16, v9
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s7, s6, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v7
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, s6
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s6, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s4, s6, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s4, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v55, v48, 16, v8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s22, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s1, s58
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v54, v6, 16, v10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s41, s3, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s43, s17, s60
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s5, s6, 0x10010
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s42, s16, s42
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s6, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s13, s6, s5
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s13, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s21, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s20, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s23, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s45, s27, s73
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[54:55]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[6:7], 24, v[66:67]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[7:8], 24, v[70:71]
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s12, 0x10010
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[18:19]
-; GFX11-TRUE16-NEXT:    s_add_i32 s11, s11, s12
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s12, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s11, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s12, s12, s11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[9:10], 24, v[16:17]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[10:11], 24, v[14:15]
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s12, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[12:13], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s44, s26, s13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 24, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 8, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v54
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 8, v54
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 24, v67
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 8, v67
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v66
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 8, v66
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 24, v71
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 8, v71
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v70
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v70
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 24, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s19, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s18, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s25, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s24, s10
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[44:45], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[8:9], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[6:7], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[42:43], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[46:47], s[40:41], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[56:57], s[28:29], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[36:37], s[10:11], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[38:39], s[4:5], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s45, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s45, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s44, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s44, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s11, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s10, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s9, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s8, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s7, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s7, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s6, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s5, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s5, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s4, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s43, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s43, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s42, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s41, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s41, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s40, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s40, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s29, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s29, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s28, 8
-; GFX11-TRUE16-NEXT:    s_branch .LBB91_5
-; GFX11-TRUE16-NEXT:  .LBB91_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr104_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr103_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr102_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr99_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr86_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr98_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr97_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr85_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr84_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr96_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr87_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 2
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 8
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 9
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 10
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s5, 12
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 13
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 14
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 15
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 17
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 18
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s5, 19
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 20
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 21
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 22
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 23
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 24
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 25
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s5, 26
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 27
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 28
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 29
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 30
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v43, s4, 31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 2
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s4, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s74, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s75, 9
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB91_2
-; GFX11-TRUE16-NEXT:  .LBB91_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s94
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s94, v43, 25
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v43, 18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, s36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, s38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v81.l, s37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v86.l, s39
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s95, v43, 26
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v43, 19
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v42, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v43, 11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s87
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s96
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s59
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s61
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s63
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s72
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s73
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s97
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s98
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s69
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, s70
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s55
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, s71
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, s48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v69.l, s80
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v70.l, s81
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, s50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v68.l, s64
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v80.l, s82
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v71.l, s51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s65
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v82.l, s83
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s84
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s66
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.l, s85
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v83.l, s52
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s67
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v96.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.l, s53
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v99.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v100.l, s86
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v97.l, s54
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v98.l, s68
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s74
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s76
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s78
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s88
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s90
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s92
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s58, v43, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s59, v43, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s60, v43, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s61, v43, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s62, v43, 21
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s63, v43, 28
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s72, v42, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s73, v42, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s13, v42, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s15, v42, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s45, v42, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s44, v42, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s47, v42, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s11, v42, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s57, v43, 31
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s10, v43, 30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s74, v43, 29
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s9, v43, 27
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s75, v43, 24
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s8, v43, 23
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s76, v43, 22
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s77, v43, 20
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s78, v43, 17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s79, v43, 16
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s88, v43, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s89, v43, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s90, v43, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s91, v43, 9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s92, s100
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s93, v43, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s43, v43, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s95, v43, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v42, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s42, v43, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s41, v43, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v43, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v43, 3
-; GFX11-TRUE16-NEXT:  .LBB91_5: ; %end
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s104
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s56
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s4
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s103
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s58
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s102
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s4
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s99
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s46
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s4
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s59
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s31
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s4
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s95
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s14
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s60
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s93
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s91
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s90
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s38
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s89
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s61
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s88
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v112, s0 :: v_dual_mov_b32 v113, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v114, s2 :: v_dual_mov_b32 v115, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s79
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s78
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s30
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s62
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s76
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s75
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s94
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s63
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s74
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[112:115], off offset:16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s57
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s36
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s72
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s47
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s44
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s45
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s92
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s15
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s73
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s13
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v100, 8, v100
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v101, s2 :: v_dual_lshlrev_b32 v12, 8, v12
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v102, s3 :: v_dual_and_b32 v99, 0xff, v99
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v87, 0xff, v87
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v99, v12
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v99, s0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v100
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v86
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v84, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v23, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v98
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v87, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v22, v86
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v27, v96
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v83
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v29, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v24, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v84, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v26, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v97
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v21, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v22, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v30
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v16, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v22, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v12, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v33
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v12, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v20, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v71
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v16, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v17, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v22, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v28, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v20, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v10, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v18, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v19, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v64
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v20, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v50
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v65
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v22, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v19, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v21, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v55
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v54
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v53
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v49
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v21, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v27, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v28, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v30, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v100, s1 :: v_dual_lshlrev_b32 v13, 16, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v12, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v19, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v13
-; GFX11-TRUE16-NEXT:    s_clause 0x5
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[99:102], off offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[23:26], off offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[14:17], off offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[18:21], off offset:112
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s104, v41, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s103, v41, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s102, v41, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s101, v41, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s100, v41, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s99, v41, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s98, v41, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s97, v41, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s96, v41, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s87, v40, 31
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s86, v40, 30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s85, v40, 29
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s84, v40, 28
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s83, v40, 27
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s82, v40, 26
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s81, v40, 25
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s80, v40, 24
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s71, v40, 23
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s70, v40, 22
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s69, v40, 21
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s68, v40, 20
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v40, 19
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v40, 18
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v40, 17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v40, 16
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v40, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v40, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v40, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v40, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v40, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v40, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v40, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v40, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v40, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v40, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v40, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v40, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v40, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v40, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v128i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:12
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v41, s96, 0
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s72, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s73, v2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v41, s97, 1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v41, s98, 2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v41, s99, 3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s46, v11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v41, s100, 4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s45, v14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v41, s101, 5
-; GFX11-FAKE16-NEXT:    s_mov_b32 vcc_hi, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v41, s102, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v41, s103, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v41, s104, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s49, 9
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s50, 10
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s51, 11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s52, 12
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s53, 13
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s54, 14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s55, 15
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s64, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s65, 17
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s66, 18
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s67, 19
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s68, 20
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s69, 21
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s70, 22
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s71, 23
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s80, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s81, 25
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s82, 26
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s83, 27
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s84, 28
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s85, 29
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s86, 30
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s87, 31
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB91_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[26:27], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 15
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s99, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s101, s1, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 14
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s0, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s85, s45, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s45, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 17
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s45, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s87, s44, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s86, s44, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 18
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s81, s47, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s98, s47, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s84, s47, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 19
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s46, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s70, s57, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s97, s57, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 13
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s80, s57, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s83, s56, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s82, s56, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 20
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s66, s59, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s59, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s69, s59, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 21
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s71, s58, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s58, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s55, s61, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 22
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s61, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s65, s61, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s68, s60, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 23
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s67, s60, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s51, s63, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s96, s63, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s54, s63, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s62, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s64, s62, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s73, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s73, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s50, s73, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 25
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s53, s72, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s52, s72, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s29, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 26
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s29, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s28, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 27
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s28, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[2:3], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 11
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[74:75], s[44:45], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[76:77], s[46:47], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 28
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[78:79], s[56:57], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[88:89], s[58:59], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[90:91], s[60:61], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 29
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[92:93], s[62:63], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[72:73], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 30
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s19, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 31
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s19, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 10
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s19, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v42, s4, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s18, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v42, s4, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s18, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v42, s4, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s17, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v42, s4, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s17, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s17, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v42, s4, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v42, s4, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s16, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v42, s4, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s3, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v42, s4, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s3, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v42, s4, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s46, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s12, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s13, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[24:25], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s12, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s13, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s12, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s13, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s12, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s13, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB91_4
-; GFX11-FAKE16-NEXT:  .LBB91_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s29, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, s47, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s1, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s15, s47, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s29, 16
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s45, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s45, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s78, s28, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s6, s47, 0x10010
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s79, s28, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s6, s47
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s73, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s77, s73, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s75, s72, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s76, s72, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, s63, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s74, s63, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s72, s62, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s73, s62, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s63, s61, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s62, s61, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s61, s60, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s60, s60, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s41, s59, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s40, s59, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s28, s58, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s29, s58, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s13, s57, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s57, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, s56, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s43, s56, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, s46, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s46, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s44, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s44, 16
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s47, 22
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s44, s47, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s44, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s78
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s1, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s79
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v4, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s1, 0x10010
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s1
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s1, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s1, s1, s45
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s44
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v4, v6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v2
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s44, 0x10010
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s44
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s44, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v6
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s44, s44, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v3, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s77
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s5, s0, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v22, 16, v4
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s5, s0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s44, 16
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s0, 22
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s0, s0, s45
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s3, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s44
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v8, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v9
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v6, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v2, 16, v3
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s44, 0x10010
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s44
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s44, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s44, s44, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s76
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s44, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s75
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s3, v10
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 24, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v6
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s3, 0x10010
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s3
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s3, 22
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v4
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s3, s3, s45
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s2, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s44
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v8, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v4
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s44, 0x10010
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v24
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s44
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s44, 22
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s74
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v25, 16, v5
-; GFX11-FAKE16-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s44, s44, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v14
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s2, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s11
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s11, s2, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s11, s2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s44, 16
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s2, 22
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s2, s2, s45
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s17, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s44
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v3, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v2, 16, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 16, v13
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v2, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s44, 0x10010
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s44
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s44, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s44, s44, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s17, s17, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s73
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s17
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s72
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s17, v4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s44, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v5, v2
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s17, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v27
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s17
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s17, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v28, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v1, 16, 1
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s17, s17, s45
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s16, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s17, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v3, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s63
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 24, v16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v3, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v5, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v29
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v8, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s44
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v8
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s44, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s44
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s44, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s44, s44, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s16, s16, 16
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s44, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s16, v8
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s62
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s16, 0x10010
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s16
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s16, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s16, s16, s45
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s19, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s44
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v5
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v4, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s16, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v2, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v10
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v9, v4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s60
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v4
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s44, 0x10010
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s61
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s44
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s44, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s44, s44, s45
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s19, s19, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s19
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v8, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s44, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s19, v10
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v9, v8
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s45, s19, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v2
-; GFX11-FAKE16-NEXT:    s_add_i32 s45, s45, s19
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s19, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s45, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s19, s19, s45
-; GFX11-FAKE16-NEXT:    s_and_b32 s44, s18, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v3, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s44
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s19, s19, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s29
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s41
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s47, s17, s72
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s44, s41, 0x10010
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    s_add_i32 s44, s44, s41
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s41, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s44, 0x7fff
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s40
-; GFX11-FAKE16-NEXT:    s_and_b32 s45, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s41, s41, s44
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s18, s18, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v30, 16, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s18, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v1, 16, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s40, s18, 0x10010
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s28
-; GFX11-FAKE16-NEXT:    s_add_i32 s44, s40, s18
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s41, 16
-; GFX11-FAKE16-NEXT:    s_addk_i32 s44, 0x7fff
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s18, 22
-; GFX11-FAKE16-NEXT:    s_and_b32 s41, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s18, s18, s44
-; GFX11-FAKE16-NEXT:    s_and_b32 s41, s21, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s41
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v9, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s18, s18, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s28, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v9
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v8, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v10, 16, 1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s29, s28, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v1
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, s28
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s28, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s29, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s41, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s28, s28, s29
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s21, s21, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s21
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s44, s2, s11
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s21, v11
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v4, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s29, s21, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, s21
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s21, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s29, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s28, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s21, s21, s29
-; GFX11-FAKE16-NEXT:    s_and_b32 s28, s20, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s28
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v10
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s21, 16
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s45, s3, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s46, s16, s46
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s28, s13, 0x10010
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v34
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, s13
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s13, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s28, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s29, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s13, s13, s28
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s20, s20, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s20
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v3, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v35
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s20, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v9
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s10, s20, 0x10010
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s10, s20
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s13, 16
-; GFX11-FAKE16-NEXT:    s_addk_i32 s28, 0x7fff
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s20, 22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v3
-; GFX11-FAKE16-NEXT:    s_and_b32 s13, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s13, s20, s28
-; GFX11-FAKE16-NEXT:    s_and_b32 s20, s23, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s42
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s20
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s43
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s28, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v19
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s20, s28, 0x10010
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v8, 16, 1
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s20, s28
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s13, 16
-; GFX11-FAKE16-NEXT:    s_addk_i32 s29, 0x7fff
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s28, 22
-; GFX11-FAKE16-NEXT:    s_and_b32 s13, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s13, s28, s29
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s23, s23, 16
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v4, v8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s13, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s23, v2
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s28, s23, 0x10010
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, s23
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s23, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s28, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s13, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s13, s23, s28
-; GFX11-FAKE16-NEXT:    s_and_b32 s23, s22, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v3, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s23
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s14
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s13, 16
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v8, 16, 1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v71, v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s12
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s15, s14, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v1
-; GFX11-FAKE16-NEXT:    s_add_i32 s15, s15, s14
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s14, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s15, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s13, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s13, s14, s15
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s14, s22, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s14
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v38
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v9, v8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s13, 16
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v10
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v70, v2, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v9
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s12, s14, 0x10010
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v8
-; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, s14
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s14, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s12, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s15, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s12, s14, s12
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, s25, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s14
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s9
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s12, 16
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s14, s9, 0x10010
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, s9
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s14, 0x7fff
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s9, s9, s14
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s12, s25, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v3, v4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s9, 16
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v8, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s12, s8, 0x10010
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v9, 16, 1
-; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, s8
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s12, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s8, s8, s12
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s8, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v2, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s9
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v12, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s0, s5
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s9, s7, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v3
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, s7
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s7, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s9, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s7, s7, s9
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s24, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s8
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v8, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s7, 16
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v10
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v3, v2, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v12, 16, 1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s4, s8, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v2
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, s8
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s4, s8, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s27, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v52, 0x40c00000, s6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s4, 16
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v9, 16, 1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s7, s6, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, s6
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s6, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s4, s6, s7
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v4, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s4, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v49
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v8
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v51
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v66, v1, 16, v11
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s7, s6, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, s6
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s6, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s4, s6, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s26, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s4, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v52
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v55, v50, 16, v4
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s22, s13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v54, v2, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v67, v48, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[17:18]
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s5, s6, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[9:10], 24, v[15:16]
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s6, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s14, s6, s5
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s26, 16
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s20, s10
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s14, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[10:11], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[6:7]
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s1, s58
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[1:2], 24, v[54:55]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[2:3], 24, v[66:67]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[70:71]
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s10, s11, 0x10010
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[4:5], 24, v[19:20]
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, s11
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s11, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s19, s60
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s10, 16
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s18, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s23, s62
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 24, v55
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 8, v55
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v54
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 8, v54
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 24, v67
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 8, v67
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v66
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 8, v66
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 24, v71
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 8, v71
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v70
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v70
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 8, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 8, v6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s21, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s25, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s57, s27, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s56, s26, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s24, s12
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[8:9], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[4:5], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[46:47], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[44:45], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[42:43], s[28:29], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 vcc, s[56:57], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[34:35], s[10:11], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[6:7], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s57, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s57, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s56, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s56, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s11, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s11, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s10, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s9, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s9, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s8, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s7, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s7, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s6, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s5, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s5, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s4, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s47, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s47, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s46, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s46, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s45, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s45, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s99, s44, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s44, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s101, s29, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s29, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s28, 8
-; GFX11-FAKE16-NEXT:    s_branch .LBB91_5
-; GFX11-FAKE16-NEXT:  .LBB91_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr104
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr103
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr102
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr101
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr100
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr99
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr96
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr97
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr98
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr86
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr87
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr85
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s5, 1
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s4, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s5, 3
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s74, 4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s75, 5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s74, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v43, s75, 7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    s_branch .LBB91_2
-; GFX11-FAKE16-NEXT:  .LBB91_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s94 :: v_dual_mov_b32 v11, s30
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s94, v43, 2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v96, s37 :: v_dual_mov_b32 v87, s34
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s49 :: v_dual_mov_b32 v7, s35
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s95, v43, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 vcc_lo, v43, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v43, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v43, 4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s44 :: v_dual_mov_b32 v51, s45
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, s10 :: v_dual_mov_b32 v49, s46
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v39, s47 :: v_dual_mov_b32 v48, s98
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s56 :: v_dual_mov_b32 v37, s97
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s57 :: v_dual_mov_b32 v35, s58
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s59 :: v_dual_mov_b32 v33, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s60 :: v_dual_mov_b32 v31, s61
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s8 :: v_dual_mov_b32 v29, s62
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s63 :: v_dual_mov_b32 v28, s96
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s72 :: v_dual_mov_b32 v25, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s73 :: v_dual_mov_b32 v23, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v21, s29 :: v_dual_mov_b32 v22, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v53, s87 :: v_dual_mov_b32 v54, s86
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s85 :: v_dual_mov_b32 v12, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v65, s4 :: v_dual_mov_b32 v66, s48
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v55, s81 :: v_dual_mov_b32 v64, s84
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v69, s83 :: v_dual_mov_b32 v70, s82
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v67, s70 :: v_dual_mov_b32 v68, s80
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v80, s71 :: v_dual_mov_b32 v19, s39
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v71, s66 :: v_dual_mov_b32 v20, s69
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, s68 :: v_dual_mov_b32 v17, s67
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v81, s55 :: v_dual_mov_b32 v18, s65
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v84, s38 :: v_dual_mov_b32 v15, s64
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v83, s51 :: v_dual_mov_b32 v16, s54
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s53 :: v_dual_mov_b32 v13, s52
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v85, s36 :: v_dual_mov_b32 v14, s50
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s74 :: v_dual_mov_b32 v2, s76
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s78 :: v_dual_mov_b32 v4, s88
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s90 :: v_dual_mov_b32 v9, s92
-; GFX11-FAKE16-NEXT:    s_mov_b32 s58, s11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s59, v43, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s72, v43, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s60, v43, 10
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s61, v43, 11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s62, v43, 12
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s63, v43, 13
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s73, v43, 14
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s13, v43, 15
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s15, v43, 16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s41, v43, 17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s43, v43, 18
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s56, v43, 19
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s11, v43, 20
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s57, v43, 21
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s10, v43, 22
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s74, v43, 23
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s9, v43, 24
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s75, v43, 25
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s8, v43, 26
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s76, v43, 27
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s77, v43, 28
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s78, v43, 29
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s79, v43, 30
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s88, v43, 31
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s89, v42, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s90, v42, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s91, v42, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s92, v42, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s47, v42, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s93, v42, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 vcc_hi, v43, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s46, v42, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v43, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s95, v42, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s45, v42, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v43, 5
-; GFX11-FAKE16-NEXT:  .LBB91_5: ; %end
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s104, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s103, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s42, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s4
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s102, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s58, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s101, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s4
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s100, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s99, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s40, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s4
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s45, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s59, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s95, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s4
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v97, s0 :: v_dual_mov_b32 v98, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v99, s2 :: v_dual_mov_b32 v100, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s46, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s93, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s14, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s47, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s72, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s92, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s91, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s90, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s12, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s19, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s89, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s60, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s88, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v112, s0 :: v_dual_mov_b32 v113, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v114, s2 :: v_dual_mov_b32 v115, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s79, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s78, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s30, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s21, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s77, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s61, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s76, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s8, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s75, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s94, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s23, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s9, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s62, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s74, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[97:100], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[112:115], off offset:16
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v97, s0 :: v_dual_mov_b32 v98, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v99, s2 :: v_dual_mov_b32 v100, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s10, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s57, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s34, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s11, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s25, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s63, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s56, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s43, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s41, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, vcc_lo, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s15, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s27, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s13, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s73, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v112, s0 :: v_dual_and_b32 v23, 0xff, v23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v113, s1 :: v_dual_lshlrev_b32 v6, 8, v6
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v114, s2 :: v_dual_lshlrev_b32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v115, s3 :: v_dual_and_b32 v96, 0xff, v96
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v23, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v96, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v6, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v22
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v87
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v86
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v11, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v22, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v26, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v24, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v25
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v85
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v84
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v83
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v22, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v24, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v25, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v26, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v6, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v11, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v13, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v15, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v32
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v31
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v30
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v81
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v35
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v11, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v13, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v17, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v80
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v34
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v33
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v71
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v70
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v14, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v17, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v19, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v21, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v68
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v67
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v49
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v65
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v39
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v64
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v10, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v16, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v18, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v19, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v48
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v55
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v52
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v54
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v53
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v51
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v50
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v16, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v18, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v20, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v21, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v22, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v3, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v10, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v11, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v17, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v12, v5
-; GFX11-FAKE16-NEXT:    s_clause 0x5
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[97:100], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[112:115], off offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[23:26], off offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[6:9], off offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s104, v41, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s103, v41, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s102, v41, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s101, v41, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s100, v41, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s99, v41, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s98, v41, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s97, v41, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s96, v41, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s87, v40, 31
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s86, v40, 30
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s85, v40, 29
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s84, v40, 28
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s83, v40, 27
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s82, v40, 26
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s81, v40, 25
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s80, v40, 24
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s71, v40, 23
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s70, v40, 22
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s69, v40, 21
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s68, v40, 20
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v40, 19
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v40, 18
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v40, 17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v40, 16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v40, 15
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v40, 14
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v40, 13
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v40, 12
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v40, 11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v40, 10
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v40, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v40, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v40, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v40, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v40, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v40, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v40, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v40, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v64bf16_to_v128i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32
+; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:12
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v41, s96, 0
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
+; GFX11-NEXT:    v_readfirstlane_b32 s72, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s73, v2
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v41, s97, 1
+; GFX11-NEXT:    v_readfirstlane_b32 s62, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s63, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s60, v5
+; GFX11-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v41, s98, 2
+; GFX11-NEXT:    v_readfirstlane_b32 s61, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s58, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s59, v8
+; GFX11-NEXT:    v_writelane_b32 v40, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v41, s99, 3
+; GFX11-NEXT:    v_readfirstlane_b32 s56, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s57, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s46, v11
+; GFX11-NEXT:    v_writelane_b32 v40, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v41, s100, 4
+; GFX11-NEXT:    v_readfirstlane_b32 s47, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s44, v13
+; GFX11-NEXT:    v_readfirstlane_b32 s45, v14
+; GFX11-NEXT:    v_writelane_b32 v40, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v41, s101, 5
+; GFX11-NEXT:    s_mov_b32 vcc_hi, 0
+; GFX11-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
+; GFX11-NEXT:    ; implicit-def: $vgpr43 : SGPR spill to VGPR lane
+; GFX11-NEXT:    ; implicit-def: $vgpr42 : SGPR spill to VGPR lane
+; GFX11-NEXT:    v_writelane_b32 v40, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v41, s102, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v41, s103, 7
+; GFX11-NEXT:    v_writelane_b32 v40, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v41, s104, 8
+; GFX11-NEXT:    v_writelane_b32 v40, s49, 9
+; GFX11-NEXT:    v_writelane_b32 v40, s50, 10
+; GFX11-NEXT:    v_writelane_b32 v40, s51, 11
+; GFX11-NEXT:    v_writelane_b32 v40, s52, 12
+; GFX11-NEXT:    v_writelane_b32 v40, s53, 13
+; GFX11-NEXT:    v_writelane_b32 v40, s54, 14
+; GFX11-NEXT:    v_writelane_b32 v40, s55, 15
+; GFX11-NEXT:    v_writelane_b32 v40, s64, 16
+; GFX11-NEXT:    v_writelane_b32 v40, s65, 17
+; GFX11-NEXT:    v_writelane_b32 v40, s66, 18
+; GFX11-NEXT:    v_writelane_b32 v40, s67, 19
+; GFX11-NEXT:    v_writelane_b32 v40, s68, 20
+; GFX11-NEXT:    v_writelane_b32 v40, s69, 21
+; GFX11-NEXT:    v_writelane_b32 v40, s70, 22
+; GFX11-NEXT:    v_writelane_b32 v40, s71, 23
+; GFX11-NEXT:    v_writelane_b32 v40, s80, 24
+; GFX11-NEXT:    v_writelane_b32 v40, s81, 25
+; GFX11-NEXT:    v_writelane_b32 v40, s82, 26
+; GFX11-NEXT:    v_writelane_b32 v40, s83, 27
+; GFX11-NEXT:    v_writelane_b32 v40, s84, 28
+; GFX11-NEXT:    v_writelane_b32 v40, s85, 29
+; GFX11-NEXT:    v_writelane_b32 v40, s86, 30
+; GFX11-NEXT:    v_writelane_b32 v40, s87, 31
+; GFX11-NEXT:    s_cbranch_scc0 .LBB91_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s4, s27, 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[26:27], 24
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 15
+; GFX11-NEXT:    s_lshr_b32 s4, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s99, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s100, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s101, s1, 24
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 14
+; GFX11-NEXT:    s_lshr_b32 s4, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s11, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s102, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s103, s0, 16
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s104, s0, 8
+; GFX11-NEXT:    s_lshr_b32 s85, s45, 24
+; GFX11-NEXT:    s_lshr_b32 s10, s45, 16
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 17
+; GFX11-NEXT:    s_lshr_b32 s4, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s5, s45, 8
+; GFX11-NEXT:    s_lshr_b32 s87, s44, 16
+; GFX11-NEXT:    s_lshr_b32 s86, s44, 8
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 18
+; GFX11-NEXT:    s_lshr_b32 s4, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s81, s47, 24
+; GFX11-NEXT:    s_lshr_b32 s98, s47, 16
+; GFX11-NEXT:    s_lshr_b32 s84, s47, 8
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 19
+; GFX11-NEXT:    s_lshr_b32 s4, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s46, 8
+; GFX11-NEXT:    s_lshr_b32 s70, s57, 24
+; GFX11-NEXT:    s_lshr_b32 s97, s57, 16
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 13
+; GFX11-NEXT:    s_lshr_b32 s4, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s80, s57, 8
+; GFX11-NEXT:    s_lshr_b32 s83, s56, 16
+; GFX11-NEXT:    s_lshr_b32 s82, s56, 8
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 20
+; GFX11-NEXT:    s_lshr_b32 s4, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s66, s59, 24
+; GFX11-NEXT:    s_lshr_b32 s9, s59, 16
+; GFX11-NEXT:    s_lshr_b32 s69, s59, 8
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 21
+; GFX11-NEXT:    s_lshr_b32 s4, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s71, s58, 16
+; GFX11-NEXT:    s_lshr_b32 s39, s58, 8
+; GFX11-NEXT:    s_lshr_b32 s55, s61, 24
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 22
+; GFX11-NEXT:    s_lshr_b32 s4, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s8, s61, 16
+; GFX11-NEXT:    s_lshr_b32 s65, s61, 8
+; GFX11-NEXT:    s_lshr_b32 s68, s60, 16
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 23
+; GFX11-NEXT:    s_lshr_b32 s4, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s67, s60, 8
+; GFX11-NEXT:    s_lshr_b32 s51, s63, 24
+; GFX11-NEXT:    s_lshr_b32 s96, s63, 16
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 12
+; GFX11-NEXT:    s_lshr_b32 s4, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s54, s63, 8
+; GFX11-NEXT:    s_lshr_b32 s38, s62, 16
+; GFX11-NEXT:    s_lshr_b32 s64, s62, 8
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 24
+; GFX11-NEXT:    s_lshr_b32 s4, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s36, s73, 24
+; GFX11-NEXT:    s_lshr_b32 s7, s73, 16
+; GFX11-NEXT:    s_lshr_b32 s50, s73, 8
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 25
+; GFX11-NEXT:    s_lshr_b32 s4, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s53, s72, 16
+; GFX11-NEXT:    s_lshr_b32 s52, s72, 8
+; GFX11-NEXT:    s_lshr_b32 s34, s29, 24
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 26
+; GFX11-NEXT:    s_lshr_b32 s4, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s6, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s35, s29, 8
+; GFX11-NEXT:    s_lshr_b32 s37, s28, 16
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 27
+; GFX11-NEXT:    s_lshr_b32 s4, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s49, s28, 8
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[2:3], 24
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 11
+; GFX11-NEXT:    s_lshr_b32 s4, s21, 8
+; GFX11-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b64 s[74:75], s[44:45], 24
+; GFX11-NEXT:    s_lshr_b64 s[76:77], s[46:47], 24
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 28
+; GFX11-NEXT:    s_lshr_b32 s4, s20, 16
+; GFX11-NEXT:    s_lshr_b64 s[78:79], s[56:57], 24
+; GFX11-NEXT:    s_lshr_b64 s[88:89], s[58:59], 24
+; GFX11-NEXT:    s_lshr_b64 s[90:91], s[60:61], 24
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 29
+; GFX11-NEXT:    s_lshr_b32 s4, s20, 8
+; GFX11-NEXT:    s_lshr_b64 s[92:93], s[62:63], 24
+; GFX11-NEXT:    s_lshr_b64 s[94:95], s[72:73], 24
+; GFX11-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 30
+; GFX11-NEXT:    s_lshr_b32 s4, s19, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 31
+; GFX11-NEXT:    s_lshr_b32 s4, s19, 16
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 10
+; GFX11-NEXT:    s_lshr_b32 s4, s19, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v42, s4, 0
+; GFX11-NEXT:    s_lshr_b32 s4, s18, 16
+; GFX11-NEXT:    v_writelane_b32 v42, s4, 1
+; GFX11-NEXT:    s_lshr_b32 s4, s18, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v42, s4, 2
+; GFX11-NEXT:    s_lshr_b32 s4, s17, 24
+; GFX11-NEXT:    v_writelane_b32 v42, s4, 3
+; GFX11-NEXT:    s_lshr_b32 s4, s17, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 9
+; GFX11-NEXT:    s_lshr_b32 s4, s17, 8
+; GFX11-NEXT:    v_writelane_b32 v42, s4, 4
+; GFX11-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v42, s4, 5
+; GFX11-NEXT:    s_lshr_b32 s4, s16, 8
+; GFX11-NEXT:    v_writelane_b32 v42, s4, 6
+; GFX11-NEXT:    s_lshr_b32 s4, s3, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v42, s4, 7
+; GFX11-NEXT:    s_lshr_b32 s4, s3, 16
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 8
+; GFX11-NEXT:    s_lshr_b32 s4, s3, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v42, s4, 8
+; GFX11-NEXT:    s_lshr_b32 s4, s46, 16
+; GFX11-NEXT:    v_writelane_b32 v43, s12, 6
+; GFX11-NEXT:    v_writelane_b32 v43, s13, 7
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[24:25], 24
+; GFX11-NEXT:    v_writelane_b32 v43, s12, 4
+; GFX11-NEXT:    v_writelane_b32 v43, s13, 5
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[22:23], 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v43, s12, 2
+; GFX11-NEXT:    v_writelane_b32 v43, s13, 3
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
+; GFX11-NEXT:    v_writelane_b32 v43, s12, 0
+; GFX11-NEXT:    v_writelane_b32 v43, s13, 1
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
+; GFX11-NEXT:    s_cbranch_vccnz .LBB91_4
+; GFX11-NEXT:  .LBB91_2: ; %cmp.true
+; GFX11-NEXT:    s_and_b32 s4, s29, 0xffff0000
+; GFX11-NEXT:    s_and_b32 s14, s47, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
+; GFX11-NEXT:    s_and_b32 s4, s1, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s15, s47, 16
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s4
+; GFX11-NEXT:    s_lshl_b32 s6, s29, 16
+; GFX11-NEXT:    v_bfe_u32 v3, v1, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s6
+; GFX11-NEXT:    s_and_b32 s8, s45, 0xffff0000
+; GFX11-NEXT:    v_readfirstlane_b32 s47, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v1
+; GFX11-NEXT:    s_lshl_b32 s7, s45, 16
+; GFX11-NEXT:    s_and_b32 s78, s28, 0xffff0000
+; GFX11-NEXT:    s_bfe_u32 s6, s47, 0x10010
+; GFX11-NEXT:    s_lshl_b32 s79, s28, 16
+; GFX11-NEXT:    s_add_i32 s45, s6, s47
+; GFX11-NEXT:    s_and_b32 s5, s73, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s77, s73, 16
+; GFX11-NEXT:    s_and_b32 s75, s72, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s76, s72, 16
+; GFX11-NEXT:    s_and_b32 s11, s63, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s74, s63, 16
+; GFX11-NEXT:    s_and_b32 s72, s62, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s73, s62, 16
+; GFX11-NEXT:    s_and_b32 s63, s61, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s62, s61, 16
+; GFX11-NEXT:    s_and_b32 s61, s60, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s60, s60, 16
+; GFX11-NEXT:    s_and_b32 s41, s59, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s40, s59, 16
+; GFX11-NEXT:    s_and_b32 s28, s58, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s29, s58, 16
+; GFX11-NEXT:    s_and_b32 s13, s57, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s10, s57, 16
+; GFX11-NEXT:    s_and_b32 s42, s56, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s43, s56, 16
+; GFX11-NEXT:    s_and_b32 s12, s46, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s9, s46, 16
+; GFX11-NEXT:    s_and_b32 s4, s44, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s6, s44, 16
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_bitset1_b32 s47, 22
+; GFX11-NEXT:    v_bfe_u32 v4, v2, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_cselect_b32 s44, s47, s45
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v2
+; GFX11-NEXT:    s_lshr_b32 s58, s44, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s1
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s78
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v3
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s79
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v4, v5, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    s_bfe_u32 s45, s1, 0x10010
+; GFX11-NEXT:    v_bfe_u32 v4, v6, 16, 1
+; GFX11-NEXT:    s_add_i32 s45, s45, s1
+; GFX11-NEXT:    s_bitset1_b32 s1, 22
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s1, s1, s45
+; GFX11-NEXT:    s_and_b32 s44, s0, 0xffff0000
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v2
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s44
+; GFX11-NEXT:    v_bfe_u32 v5, v7, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v4, v6
+; GFX11-NEXT:    s_lshr_b32 s1, s1, 16
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s44, v2
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    s_bfe_u32 s45, s44, 0x10010
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_add_i32 s45, s45, s44
+; GFX11-NEXT:    s_bitset1_b32 s44, 22
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v5, v7
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v6
+; GFX11-NEXT:    s_cselect_b32 s44, s44, s45
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v3, v5, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v21
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s5
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v8, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s77
+; GFX11-NEXT:    s_bfe_u32 s5, s0, 0x10010
+; GFX11-NEXT:    v_lshl_or_b32 v7, v22, 16, v4
+; GFX11-NEXT:    s_add_i32 s45, s5, s0
+; GFX11-NEXT:    s_lshr_b32 s5, s44, 16
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_bitset1_b32 s0, 22
+; GFX11-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s0, s0, s45
+; GFX11-NEXT:    s_and_b32 s44, s3, 0xffff0000
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v1
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s44
+; GFX11-NEXT:    v_bfe_u32 v6, v8, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v1, v5, 16, 1
+; GFX11-NEXT:    s_lshr_b32 s0, s0, 16
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v23
+; GFX11-NEXT:    v_readfirstlane_b32 s44, v9
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v6, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v5
+; GFX11-NEXT:    v_lshl_or_b32 v6, v2, 16, v3
+; GFX11-NEXT:    s_bfe_u32 s45, s44, 0x10010
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v5
+; GFX11-NEXT:    s_add_i32 s45, s45, s44
+; GFX11-NEXT:    s_bitset1_b32 s44, 22
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v8
+; GFX11-NEXT:    s_cselect_b32 s44, s44, s45
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s3
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s76
+; GFX11-NEXT:    s_lshr_b32 s59, s44, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s75
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v10
+; GFX11-NEXT:    v_bfe_u32 v8, v9, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 24, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 16, v6
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v2, v4, 16, 1
+; GFX11-NEXT:    s_bfe_u32 s45, s3, 0x10010
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    s_add_i32 s45, s45, s3
+; GFX11-NEXT:    s_bitset1_b32 s3, 22
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v4
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s3, s3, s45
+; GFX11-NEXT:    s_and_b32 s44, s2, 0xffff0000
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v1
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s44
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v8, v9
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s44, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    s_lshr_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v4
+; GFX11-NEXT:    s_bfe_u32 s45, s44, 0x10010
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v24
+; GFX11-NEXT:    s_add_i32 s45, s45, s44
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v8, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_bitset1_b32 s44, 22
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s74
+; GFX11-NEXT:    v_lshl_or_b32 v14, v25, 16, v5
+; GFX11-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s44, s44, s45
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s2
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 24, v14
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_readfirstlane_b32 s2, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s11
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    s_bfe_u32 s11, s2, 0x10010
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_add_i32 s45, s11, s2
+; GFX11-NEXT:    s_lshr_b32 s11, s44, 16
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_bitset1_b32 s2, 22
+; GFX11-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s2, s2, s45
+; GFX11-NEXT:    s_and_b32 s44, s17, 0xffff0000
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v26
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s44
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_bfe_u32 v10, v3, 16, 1
+; GFX11-NEXT:    s_lshr_b32 s2, s2, 16
+; GFX11-NEXT:    v_lshl_or_b32 v13, v2, 16, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s44, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v10, v3
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 16, v13
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v2, v4, vcc_lo
+; GFX11-NEXT:    s_bfe_u32 s45, s44, 0x10010
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    s_add_i32 s45, s45, s44
+; GFX11-NEXT:    s_bitset1_b32 s44, 22
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
+; GFX11-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s44, s44, s45
+; GFX11-NEXT:    s_lshl_b32 s17, s17, 16
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s73
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s17
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v1
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s72
+; GFX11-NEXT:    v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT:    v_readfirstlane_b32 s17, v4
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    s_lshr_b32 s72, s44, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v5, v2
+; GFX11-NEXT:    s_bfe_u32 s45, s17, 0x10010
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v27
+; GFX11-NEXT:    s_add_i32 s45, s45, s17
+; GFX11-NEXT:    s_bitset1_b32 s17, 22
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
+; GFX11-NEXT:    v_lshl_or_b32 v16, v28, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v2
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_bfe_u32 v8, v1, 16, 1
+; GFX11-NEXT:    s_cselect_b32 s17, s17, s45
+; GFX11-NEXT:    s_and_b32 s44, s16, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s17, s17, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v3, v5, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s63
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 24, v16
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v2
+; GFX11-NEXT:    v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v5, v3
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v29
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v8, v1
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s44
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s44, v8
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    s_bfe_u32 s45, s44, 0x10010
+; GFX11-NEXT:    s_add_i32 s45, s45, s44
+; GFX11-NEXT:    s_bitset1_b32 s44, 22
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_and_b32 s46, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s44, s44, s45
+; GFX11-NEXT:    s_lshl_b32 s16, s16, 16
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s16
+; GFX11-NEXT:    s_lshr_b32 s46, s44, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v4, v9, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT:    v_readfirstlane_b32 s16, v8
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v3
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s62
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    s_bfe_u32 s45, s16, 0x10010
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_add_i32 s45, s45, s16
+; GFX11-NEXT:    s_bitset1_b32 s16, 22
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s16, s16, s45
+; GFX11-NEXT:    s_and_b32 s44, s19, 0xffff0000
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s44
+; GFX11-NEXT:    v_lshl_or_b32 v15, v1, 16, v5
+; GFX11-NEXT:    v_bfe_u32 v9, v4, 16, 1
+; GFX11-NEXT:    s_lshr_b32 s16, s16, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v2, v8, vcc_lo
+; GFX11-NEXT:    v_readfirstlane_b32 s44, v10
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, v9, v4
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s60
+; GFX11-NEXT:    v_or_b32_e32 v3, 0x400000, v4
+; GFX11-NEXT:    s_bfe_u32 s45, s44, 0x10010
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s61
+; GFX11-NEXT:    s_add_i32 s45, s45, s44
+; GFX11-NEXT:    s_bitset1_b32 s44, 22
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s44, s44, s45
+; GFX11-NEXT:    s_lshl_b32 s19, s19, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v9
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s19
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_bfe_u32 v9, v8, 16, 1
+; GFX11-NEXT:    s_lshr_b32 s60, s44, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s19, v10
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    v_bfe_u32 v3, v5, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v9, v8
+; GFX11-NEXT:    s_bfe_u32 s45, s19, 0x10010
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v2
+; GFX11-NEXT:    s_add_i32 s45, s45, s19
+; GFX11-NEXT:    s_bitset1_b32 s19, 22
+; GFX11-NEXT:    s_addk_i32 s45, 0x7fff
+; GFX11-NEXT:    s_and_b32 s44, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s19, s19, s45
+; GFX11-NEXT:    s_and_b32 s44, s18, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v3, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
+; GFX11-NEXT:    v_or_b32_e32 v3, 0x400000, v8
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s44
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT:    s_lshr_b32 s19, s19, 16
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s29
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s41
+; GFX11-NEXT:    v_readfirstlane_b32 s41, v4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s47, s17, s72
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v2
+; GFX11-NEXT:    v_bfe_u32 v2, v3, 16, 1
+; GFX11-NEXT:    s_bfe_u32 s44, s41, 0x10010
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    s_add_i32 s44, s44, s41
+; GFX11-NEXT:    s_bitset1_b32 s41, 22
+; GFX11-NEXT:    s_addk_i32 s44, 0x7fff
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s40
+; GFX11-NEXT:    s_and_b32 s45, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s41, s41, s44
+; GFX11-NEXT:    s_lshl_b32 s18, s18, 16
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v31
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v32
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v3
+; GFX11-NEXT:    v_lshl_or_b32 v18, v30, 16, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s18, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_lshl_or_b32 v17, v1, 16, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v2
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v3
+; GFX11-NEXT:    s_bfe_u32 s40, s18, 0x10010
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s28
+; GFX11-NEXT:    s_add_i32 s44, s40, s18
+; GFX11-NEXT:    s_lshr_b32 s40, s41, 16
+; GFX11-NEXT:    s_addk_i32 s44, 0x7fff
+; GFX11-NEXT:    s_bitset1_b32 s18, 22
+; GFX11-NEXT:    s_and_b32 s41, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s18, s18, s44
+; GFX11-NEXT:    s_and_b32 s41, s21, 0xffff0000
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s41
+; GFX11-NEXT:    v_bfe_u32 v2, v9, 16, 1
+; GFX11-NEXT:    s_lshr_b32 s18, s18, 16
+; GFX11-NEXT:    v_or_b32_e32 v3, 0x400000, v9
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
+; GFX11-NEXT:    v_readfirstlane_b32 s28, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v9
+; GFX11-NEXT:    v_bfe_u32 v4, v8, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v5, v10, 16, 1
+; GFX11-NEXT:    s_bfe_u32 s29, s28, 0x10010
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v1
+; GFX11-NEXT:    s_add_i32 s29, s29, s28
+; GFX11-NEXT:    s_bitset1_b32 s28, 22
+; GFX11-NEXT:    s_addk_i32 s29, 0x7fff
+; GFX11-NEXT:    s_and_b32 s41, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s28, s28, s29
+; GFX11-NEXT:    s_lshl_b32 s21, s21, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_add_f32_e64 v11, 0x40c00000, s21
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    s_lshr_b32 s61, s28, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v5, v10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s44, s2, s11
+; GFX11-NEXT:    v_readfirstlane_b32 s21, v11
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v4, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    s_bfe_u32 s29, s21, 0x10010
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
+; GFX11-NEXT:    s_add_i32 s29, s29, s21
+; GFX11-NEXT:    s_bitset1_b32 s21, 22
+; GFX11-NEXT:    s_addk_i32 s29, 0x7fff
+; GFX11-NEXT:    s_and_b32 s28, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s21, s21, s29
+; GFX11-NEXT:    s_and_b32 s28, s20, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
+; GFX11-NEXT:    v_or_b32_e32 v3, 0x400000, v8
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s28
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v10
+; GFX11-NEXT:    s_lshr_b32 s21, s21, 16
+; GFX11-NEXT:    s_pack_ll_b32_b16 s45, s3, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s46, s16, s46
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s13
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 24, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    s_bfe_u32 s28, s13, 0x10010
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v34
+; GFX11-NEXT:    s_add_i32 s28, s28, s13
+; GFX11-NEXT:    s_bitset1_b32 s13, 22
+; GFX11-NEXT:    s_addk_i32 s28, 0x7fff
+; GFX11-NEXT:    s_and_b32 s29, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s13, s13, s28
+; GFX11-NEXT:    s_lshl_b32 s20, s20, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s20
+; GFX11-NEXT:    v_bfe_u32 v1, v3, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s10
+; GFX11-NEXT:    v_lshl_or_b32 v20, v33, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v35
+; GFX11-NEXT:    v_readfirstlane_b32 s20, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_bfe_u32 v4, v5, 16, 1
+; GFX11-NEXT:    v_lshl_or_b32 v19, v2, 16, v9
+; GFX11-NEXT:    s_bfe_u32 s10, s20, 0x10010
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    s_add_i32 s28, s10, s20
+; GFX11-NEXT:    s_lshr_b32 s10, s13, 16
+; GFX11-NEXT:    s_addk_i32 s28, 0x7fff
+; GFX11-NEXT:    s_bitset1_b32 s20, 22
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v3
+; GFX11-NEXT:    s_and_b32 s13, vcc_lo, exec_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    s_cselect_b32 s13, s20, s28
+; GFX11-NEXT:    s_and_b32 s20, s23, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v5
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s42
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s20
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v5
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s43
+; GFX11-NEXT:    v_readfirstlane_b32 s28, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v19
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    s_bfe_u32 s20, s28, 0x10010
+; GFX11-NEXT:    v_bfe_u32 v4, v8, 16, 1
+; GFX11-NEXT:    s_add_i32 s29, s20, s28
+; GFX11-NEXT:    s_lshr_b32 s20, s13, 16
+; GFX11-NEXT:    s_addk_i32 s29, 0x7fff
+; GFX11-NEXT:    s_bitset1_b32 s28, 22
+; GFX11-NEXT:    s_and_b32 s13, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s13, s28, s29
+; GFX11-NEXT:    s_lshl_b32 s23, s23, 16
+; GFX11-NEXT:    v_bfe_u32 v5, v9, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v4, v8
+; GFX11-NEXT:    s_lshr_b32 s62, s13, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v5, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s23, v2
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    s_bfe_u32 s28, s23, 0x10010
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v9
+; GFX11-NEXT:    s_add_i32 s28, s28, s23
+; GFX11-NEXT:    s_bitset1_b32 s23, 22
+; GFX11-NEXT:    s_addk_i32 s28, 0x7fff
+; GFX11-NEXT:    s_and_b32 s13, vcc_lo, exec_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    s_cselect_b32 s13, s23, s28
+; GFX11-NEXT:    s_and_b32 s23, s22, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s15
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v36
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v3, v5, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s23
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s14
+; GFX11-NEXT:    s_lshr_b32 s23, s13, 16
+; GFX11-NEXT:    v_bfe_u32 v9, v8, 16, 1
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v10, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_lshl_or_b32 v71, v37, 16, v4
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s12
+; GFX11-NEXT:    s_bfe_u32 s15, s14, 0x10010
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v1
+; GFX11-NEXT:    s_add_i32 s15, s15, s14
+; GFX11-NEXT:    s_bitset1_b32 s14, 22
+; GFX11-NEXT:    s_addk_i32 s15, 0x7fff
+; GFX11-NEXT:    s_and_b32 s13, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s13, s14, s15
+; GFX11-NEXT:    s_lshl_b32 s14, s22, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s14
+; GFX11-NEXT:    v_bfe_u32 v1, v5, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v38
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, v9, v8
+; GFX11-NEXT:    s_lshr_b32 s13, s13, 16
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v10
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v5
+; GFX11-NEXT:    v_lshl_or_b32 v70, v2, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v9
+; GFX11-NEXT:    s_bfe_u32 s12, s14, 0x10010
+; GFX11-NEXT:    v_or_b32_e32 v3, 0x400000, v8
+; GFX11-NEXT:    s_add_i32 s12, s12, s14
+; GFX11-NEXT:    s_bitset1_b32 s14, 22
+; GFX11-NEXT:    s_addk_i32 s12, 0x7fff
+; GFX11-NEXT:    s_and_b32 s15, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s12, s14, s12
+; GFX11-NEXT:    s_and_b32 s14, s25, 0xffff0000
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s14
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s9
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    s_lshr_b32 s22, s12, 16
+; GFX11-NEXT:    v_bfe_u32 v3, v4, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
+; GFX11-NEXT:    s_bfe_u32 s14, s9, 0x10010
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    s_add_i32 s14, s14, s9
+; GFX11-NEXT:    s_bitset1_b32 s9, 22
+; GFX11-NEXT:    s_addk_i32 s14, 0x7fff
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s9, s9, s14
+; GFX11-NEXT:    s_lshl_b32 s12, s25, 16
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s8
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s12
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v3, v4
+; GFX11-NEXT:    s_lshr_b32 s63, s9, 16
+; GFX11-NEXT:    v_bfe_u32 v3, v8, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v8
+; GFX11-NEXT:    v_or_b32_e32 v11, 0x400000, v8
+; GFX11-NEXT:    s_bfe_u32 s12, s8, 0x10010
+; GFX11-NEXT:    v_bfe_u32 v12, v9, 16, 1
+; GFX11-NEXT:    s_add_i32 s12, s12, s8
+; GFX11-NEXT:    s_bitset1_b32 s8, 22
+; GFX11-NEXT:    s_addk_i32 s12, 0x7fff
+; GFX11-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    s_cselect_b32 s8, s8, s12
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    s_lshr_b32 s25, s8, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v2, v10, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s9
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s7
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v12, v9
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s6
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v11, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s0, s5
+; GFX11-NEXT:    s_bfe_u32 s9, s7, 0x10010
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v3
+; GFX11-NEXT:    s_add_i32 s9, s9, s7
+; GFX11-NEXT:    s_bitset1_b32 s7, 22
+; GFX11-NEXT:    s_addk_i32 s9, 0x7fff
+; GFX11-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s7, s7, s9
+; GFX11-NEXT:    s_lshl_b32 s8, s24, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v4
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s8
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v4, v8, 16, 1
+; GFX11-NEXT:    s_lshr_b32 s12, s7, 16
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v10
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v3, v2, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v8
+; GFX11-NEXT:    v_bfe_u32 v10, v12, 16, 1
+; GFX11-NEXT:    s_bfe_u32 s4, s8, 0x10010
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v2
+; GFX11-NEXT:    s_add_i32 s4, s4, s8
+; GFX11-NEXT:    s_bitset1_b32 s8, 22
+; GFX11-NEXT:    s_addk_i32 s4, 0x7fff
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s4, s8, s4
+; GFX11-NEXT:    s_and_b32 s6, s27, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v4
+; GFX11-NEXT:    v_add_f32_e64 v52, 0x40c00000, s6
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v8
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v10, v12
+; GFX11-NEXT:    s_lshr_b32 s24, s4, 16
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v52
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v9
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
+; GFX11-NEXT:    v_bfe_u32 v4, v9, 16, 1
+; GFX11-NEXT:    s_bfe_u32 s7, s6, 0x10010
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    s_add_i32 s7, s7, s6
+; GFX11-NEXT:    s_bitset1_b32 s6, 22
+; GFX11-NEXT:    s_addk_i32 s7, 0x7fff
+; GFX11-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s4, s6, s7
+; GFX11-NEXT:    s_lshl_b32 s6, s27, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v4, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v12
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    s_lshr_b32 s73, s4, 16
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v49
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v8
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v51
+; GFX11-NEXT:    v_lshl_or_b32 v66, v1, 16, v11
+; GFX11-NEXT:    s_bfe_u32 s7, s6, 0x10010
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v10, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    s_add_i32 s7, s7, s6
+; GFX11-NEXT:    s_bitset1_b32 s6, 22
+; GFX11-NEXT:    s_addk_i32 s7, 0x7fff
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s4, s6, s7
+; GFX11-NEXT:    s_and_b32 s6, s26, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s27, s4, 16
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s6
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v52
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v39
+; GFX11-NEXT:    v_lshl_or_b32 v55, v50, 16, v4
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s22, s13
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_lshl_or_b32 v54, v2, 16, v8
+; GFX11-NEXT:    v_lshl_or_b32 v67, v48, 16, v5
+; GFX11-NEXT:    v_lshrrev_b64 v[8:9], 24, v[17:18]
+; GFX11-NEXT:    s_bfe_u32 s5, s6, 0x10010
+; GFX11-NEXT:    v_lshrrev_b64 v[9:10], 24, v[15:16]
+; GFX11-NEXT:    s_add_i32 s5, s5, s6
+; GFX11-NEXT:    s_bitset1_b32 s6, 22
+; GFX11-NEXT:    s_addk_i32 s5, 0x7fff
+; GFX11-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s14, s6, s5
+; GFX11-NEXT:    s_lshl_b32 s4, s26, 16
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s20, s10
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
+; GFX11-NEXT:    s_lshr_b32 s13, s14, 16
+; GFX11-NEXT:    v_lshrrev_b64 v[10:11], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[6:7]
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s1, s58
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_lshrrev_b64 v[1:2], 24, v[54:55]
+; GFX11-NEXT:    v_lshrrev_b64 v[2:3], 24, v[66:67]
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[70:71]
+; GFX11-NEXT:    s_bfe_u32 s10, s11, 0x10010
+; GFX11-NEXT:    v_lshrrev_b64 v[4:5], 24, v[19:20]
+; GFX11-NEXT:    s_add_i32 s10, s10, s11
+; GFX11-NEXT:    s_bitset1_b32 s11, 22
+; GFX11-NEXT:    s_addk_i32 s10, 0x7fff
+; GFX11-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s10, s11, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s19, s60
+; GFX11-NEXT:    s_lshr_b32 s26, s10, 16
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s18, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s23, s62
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 24, v55
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 8, v55
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v54
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 8, v54
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 24, v67
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 8, v67
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v66
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 8, v66
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 24, v71
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 8, v71
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v70
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 8, v70
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 24, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 8, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 8, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 8, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 8, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 8, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 8, v6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s21, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s25, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s57, s27, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s56, s26, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s24, s12
+; GFX11-NEXT:    s_lshr_b64 s[94:95], s[8:9], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[4:5], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[46:47], 24
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[44:45], 24
+; GFX11-NEXT:    s_lshr_b64 s[42:43], s[28:29], 24
+; GFX11-NEXT:    s_lshr_b64 vcc, s[56:57], 24
+; GFX11-NEXT:    s_lshr_b64 s[34:35], s[10:11], 24
+; GFX11-NEXT:    s_lshr_b64 s[30:31], s[6:7], 24
+; GFX11-NEXT:    s_lshr_b32 s13, s57, 24
+; GFX11-NEXT:    s_lshr_b32 s15, s57, 8
+; GFX11-NEXT:    s_lshr_b32 s41, s56, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s56, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s11, 24
+; GFX11-NEXT:    s_lshr_b32 s11, s11, 8
+; GFX11-NEXT:    s_lshr_b32 s57, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s10, 8
+; GFX11-NEXT:    s_lshr_b32 s74, s9, 24
+; GFX11-NEXT:    s_lshr_b32 s9, s9, 8
+; GFX11-NEXT:    s_lshr_b32 s75, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s8, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s7, 24
+; GFX11-NEXT:    s_lshr_b32 s77, s7, 8
+; GFX11-NEXT:    s_lshr_b32 s78, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s6, 8
+; GFX11-NEXT:    s_lshr_b32 s88, s5, 24
+; GFX11-NEXT:    s_lshr_b32 s89, s5, 8
+; GFX11-NEXT:    s_lshr_b32 s90, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s4, 8
+; GFX11-NEXT:    s_lshr_b32 s92, s47, 24
+; GFX11-NEXT:    s_lshr_b32 s47, s47, 8
+; GFX11-NEXT:    s_lshr_b32 s93, s46, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s46, 8
+; GFX11-NEXT:    s_lshr_b32 s95, s45, 24
+; GFX11-NEXT:    s_lshr_b32 s45, s45, 8
+; GFX11-NEXT:    s_lshr_b32 s99, s44, 16
+; GFX11-NEXT:    s_lshr_b32 s100, s44, 8
+; GFX11-NEXT:    s_lshr_b32 s101, s29, 24
+; GFX11-NEXT:    s_lshr_b32 s102, s29, 8
+; GFX11-NEXT:    s_lshr_b32 s103, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s104, s28, 8
+; GFX11-NEXT:    s_branch .LBB91_5
+; GFX11-NEXT:  .LBB91_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr104
+; GFX11-NEXT:    ; implicit-def: $sgpr103
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr102
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr101
+; GFX11-NEXT:    ; implicit-def: $sgpr100
+; GFX11-NEXT:    ; implicit-def: $sgpr99
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr52
+; GFX11-NEXT:    ; implicit-def: $sgpr53
+; GFX11-NEXT:    ; implicit-def: $sgpr50
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr64
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr54
+; GFX11-NEXT:    ; implicit-def: $sgpr96
+; GFX11-NEXT:    ; implicit-def: $sgpr51
+; GFX11-NEXT:    ; implicit-def: $sgpr67
+; GFX11-NEXT:    ; implicit-def: $sgpr68
+; GFX11-NEXT:    ; implicit-def: $sgpr65
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr55
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr71
+; GFX11-NEXT:    ; implicit-def: $sgpr69
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr66
+; GFX11-NEXT:    ; implicit-def: $sgpr82
+; GFX11-NEXT:    ; implicit-def: $sgpr83
+; GFX11-NEXT:    ; implicit-def: $sgpr80
+; GFX11-NEXT:    ; implicit-def: $sgpr97
+; GFX11-NEXT:    ; implicit-def: $sgpr70
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr84
+; GFX11-NEXT:    ; implicit-def: $sgpr98
+; GFX11-NEXT:    ; implicit-def: $sgpr81
+; GFX11-NEXT:    ; implicit-def: $sgpr86
+; GFX11-NEXT:    ; implicit-def: $sgpr87
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr85
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 0
+; GFX11-NEXT:    v_writelane_b32 v43, s5, 1
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    v_writelane_b32 v43, s4, 2
+; GFX11-NEXT:    v_writelane_b32 v43, s5, 3
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    v_writelane_b32 v43, s74, 4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    v_writelane_b32 v43, s75, 5
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; kill: killed $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    v_writelane_b32 v43, s74, 6
+; GFX11-NEXT:    v_writelane_b32 v43, s75, 7
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    s_branch .LBB91_2
+; GFX11-NEXT:  .LBB91_4:
+; GFX11-NEXT:    v_dual_mov_b32 v10, s94 :: v_dual_mov_b32 v11, s30
+; GFX11-NEXT:    v_readlane_b32 s94, v43, 2
+; GFX11-NEXT:    v_dual_mov_b32 v96, s37 :: v_dual_mov_b32 v87, s34
+; GFX11-NEXT:    v_dual_mov_b32 v6, s49 :: v_dual_mov_b32 v7, s35
+; GFX11-NEXT:    v_readlane_b32 s95, v43, 3
+; GFX11-NEXT:    v_readlane_b32 vcc_lo, v43, 6
+; GFX11-NEXT:    v_readlane_b32 s30, v43, 0
+; GFX11-NEXT:    v_readlane_b32 s34, v43, 4
+; GFX11-NEXT:    v_dual_mov_b32 v52, s44 :: v_dual_mov_b32 v51, s45
+; GFX11-NEXT:    v_dual_mov_b32 v50, s10 :: v_dual_mov_b32 v49, s46
+; GFX11-NEXT:    v_dual_mov_b32 v39, s47 :: v_dual_mov_b32 v48, s98
+; GFX11-NEXT:    v_dual_mov_b32 v38, s56 :: v_dual_mov_b32 v37, s97
+; GFX11-NEXT:    v_dual_mov_b32 v36, s57 :: v_dual_mov_b32 v35, s58
+; GFX11-NEXT:    v_dual_mov_b32 v34, s59 :: v_dual_mov_b32 v33, s9
+; GFX11-NEXT:    v_dual_mov_b32 v32, s60 :: v_dual_mov_b32 v31, s61
+; GFX11-NEXT:    v_dual_mov_b32 v30, s8 :: v_dual_mov_b32 v29, s62
+; GFX11-NEXT:    v_dual_mov_b32 v27, s63 :: v_dual_mov_b32 v28, s96
+; GFX11-NEXT:    v_dual_mov_b32 v26, s72 :: v_dual_mov_b32 v25, s7
+; GFX11-NEXT:    v_dual_mov_b32 v24, s73 :: v_dual_mov_b32 v23, s28
+; GFX11-NEXT:    v_dual_mov_b32 v21, s29 :: v_dual_mov_b32 v22, s6
+; GFX11-NEXT:    v_dual_mov_b32 v53, s87 :: v_dual_mov_b32 v54, s86
+; GFX11-NEXT:    v_dual_mov_b32 v5, s85 :: v_dual_mov_b32 v12, s5
+; GFX11-NEXT:    v_dual_mov_b32 v65, s4 :: v_dual_mov_b32 v66, s48
+; GFX11-NEXT:    v_dual_mov_b32 v55, s81 :: v_dual_mov_b32 v64, s84
+; GFX11-NEXT:    v_dual_mov_b32 v69, s83 :: v_dual_mov_b32 v70, s82
+; GFX11-NEXT:    v_dual_mov_b32 v67, s70 :: v_dual_mov_b32 v68, s80
+; GFX11-NEXT:    v_dual_mov_b32 v80, s71 :: v_dual_mov_b32 v19, s39
+; GFX11-NEXT:    v_dual_mov_b32 v71, s66 :: v_dual_mov_b32 v20, s69
+; GFX11-NEXT:    v_dual_mov_b32 v82, s68 :: v_dual_mov_b32 v17, s67
+; GFX11-NEXT:    v_dual_mov_b32 v81, s55 :: v_dual_mov_b32 v18, s65
+; GFX11-NEXT:    v_dual_mov_b32 v84, s38 :: v_dual_mov_b32 v15, s64
+; GFX11-NEXT:    v_dual_mov_b32 v83, s51 :: v_dual_mov_b32 v16, s54
+; GFX11-NEXT:    v_dual_mov_b32 v86, s53 :: v_dual_mov_b32 v13, s52
+; GFX11-NEXT:    v_dual_mov_b32 v85, s36 :: v_dual_mov_b32 v14, s50
+; GFX11-NEXT:    v_dual_mov_b32 v1, s74 :: v_dual_mov_b32 v2, s76
+; GFX11-NEXT:    v_dual_mov_b32 v3, s78 :: v_dual_mov_b32 v4, s88
+; GFX11-NEXT:    v_dual_mov_b32 v8, s90 :: v_dual_mov_b32 v9, s92
+; GFX11-NEXT:    s_mov_b32 s58, s11
+; GFX11-NEXT:    v_readlane_b32 s59, v43, 8
+; GFX11-NEXT:    v_readlane_b32 s72, v43, 9
+; GFX11-NEXT:    v_readlane_b32 s60, v43, 10
+; GFX11-NEXT:    v_readlane_b32 s61, v43, 11
+; GFX11-NEXT:    v_readlane_b32 s62, v43, 12
+; GFX11-NEXT:    v_readlane_b32 s63, v43, 13
+; GFX11-NEXT:    v_readlane_b32 s73, v43, 14
+; GFX11-NEXT:    v_readlane_b32 s13, v43, 15
+; GFX11-NEXT:    v_readlane_b32 s15, v43, 16
+; GFX11-NEXT:    v_readlane_b32 s41, v43, 17
+; GFX11-NEXT:    v_readlane_b32 s43, v43, 18
+; GFX11-NEXT:    v_readlane_b32 s56, v43, 19
+; GFX11-NEXT:    v_readlane_b32 s11, v43, 20
+; GFX11-NEXT:    v_readlane_b32 s57, v43, 21
+; GFX11-NEXT:    v_readlane_b32 s10, v43, 22
+; GFX11-NEXT:    v_readlane_b32 s74, v43, 23
+; GFX11-NEXT:    v_readlane_b32 s9, v43, 24
+; GFX11-NEXT:    v_readlane_b32 s75, v43, 25
+; GFX11-NEXT:    v_readlane_b32 s8, v43, 26
+; GFX11-NEXT:    v_readlane_b32 s76, v43, 27
+; GFX11-NEXT:    v_readlane_b32 s77, v43, 28
+; GFX11-NEXT:    v_readlane_b32 s78, v43, 29
+; GFX11-NEXT:    v_readlane_b32 s79, v43, 30
+; GFX11-NEXT:    v_readlane_b32 s88, v43, 31
+; GFX11-NEXT:    v_readlane_b32 s89, v42, 0
+; GFX11-NEXT:    v_readlane_b32 s90, v42, 1
+; GFX11-NEXT:    v_readlane_b32 s91, v42, 2
+; GFX11-NEXT:    v_readlane_b32 s92, v42, 3
+; GFX11-NEXT:    v_readlane_b32 s47, v42, 4
+; GFX11-NEXT:    v_readlane_b32 s93, v42, 5
+; GFX11-NEXT:    v_readlane_b32 vcc_hi, v43, 7
+; GFX11-NEXT:    v_readlane_b32 s46, v42, 6
+; GFX11-NEXT:    v_readlane_b32 s31, v43, 1
+; GFX11-NEXT:    v_readlane_b32 s95, v42, 7
+; GFX11-NEXT:    v_readlane_b32 s45, v42, 8
+; GFX11-NEXT:    v_readlane_b32 s35, v43, 5
+; GFX11-NEXT:  .LBB91_5: ; %end
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s104, 8
+; GFX11-NEXT:    s_and_b32 s5, s103, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s42, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s4
+; GFX11-NEXT:    s_or_b32 s4, s5, s6
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s102, 8
+; GFX11-NEXT:    s_and_b32 s6, s58, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s101, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s5
+; GFX11-NEXT:    s_or_b32 s5, s6, s7
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 16
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s4
+; GFX11-NEXT:    s_or_b32 s1, s1, s5
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s100, 8
+; GFX11-NEXT:    s_and_b32 s5, s99, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s40, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s4
+; GFX11-NEXT:    s_or_b32 s4, s5, s6
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s45, 8
+; GFX11-NEXT:    s_and_b32 s6, s59, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s95, 8
+; GFX11-NEXT:    s_or_b32 s3, s3, s5
+; GFX11-NEXT:    s_or_b32 s5, s6, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 16
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s4
+; GFX11-NEXT:    s_or_b32 s3, s3, s5
+; GFX11-NEXT:    v_dual_mov_b32 v97, s0 :: v_dual_mov_b32 v98, s1
+; GFX11-NEXT:    v_dual_mov_b32 v99, s2 :: v_dual_mov_b32 v100, s3
+; GFX11-NEXT:    s_and_b32 s0, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s46, 8
+; GFX11-NEXT:    s_and_b32 s2, s93, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s14, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s17, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s47, 8
+; GFX11-NEXT:    s_and_b32 s4, s72, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s92, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s91, 8
+; GFX11-NEXT:    s_and_b32 s4, s90, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s12, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    s_and_b32 s4, s19, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s89, 8
+; GFX11-NEXT:    s_and_b32 s6, s60, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s88, 8
+; GFX11-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-NEXT:    s_or_b32 s5, s6, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s4, s4, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    v_dual_mov_b32 v112, s0 :: v_dual_mov_b32 v113, s1
+; GFX11-NEXT:    v_dual_mov_b32 v114, s2 :: v_dual_mov_b32 v115, s3
+; GFX11-NEXT:    s_and_b32 s0, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s79, 8
+; GFX11-NEXT:    s_and_b32 s2, s78, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s30, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s21, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s77, 8
+; GFX11-NEXT:    s_and_b32 s4, s61, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s76, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s8, 8
+; GFX11-NEXT:    s_and_b32 s4, s75, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s94, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    s_and_b32 s4, s23, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s9, 8
+; GFX11-NEXT:    s_and_b32 s6, s62, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s74, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-NEXT:    s_or_b32 s5, s6, s7
+; GFX11-NEXT:    s_and_b32 s4, s4, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    scratch_store_b128 v0, v[97:100], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[112:115], off offset:16
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    v_dual_mov_b32 v97, s0 :: v_dual_mov_b32 v98, s1
+; GFX11-NEXT:    v_dual_mov_b32 v99, s2 :: v_dual_mov_b32 v100, s3
+; GFX11-NEXT:    s_and_b32 s0, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s10, 8
+; GFX11-NEXT:    s_and_b32 s2, s57, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s34, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s4
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_lshl_b32 s2, s11, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s25, 0xff
+; GFX11-NEXT:    s_and_b32 s3, s63, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s56, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_or_b32 s2, s3, s4
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_and_b32 s3, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s43, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_or_b32 s2, s3, s4
+; GFX11-NEXT:    s_and_b32 s3, s41, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, vcc_lo, 8
+; GFX11-NEXT:    s_lshl_b32 s5, s15, 8
+; GFX11-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-NEXT:    s_and_b32 s4, s27, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s13, 8
+; GFX11-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-NEXT:    s_and_b32 s5, s73, 0xff
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s4, s4, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    v_dual_mov_b32 v112, s0 :: v_dual_and_b32 v23, 0xff, v23
+; GFX11-NEXT:    v_dual_mov_b32 v113, s1 :: v_dual_lshlrev_b32 v6, 8, v6
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v114, s2 :: v_dual_lshlrev_b32 v11, 8, v11
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    v_dual_mov_b32 v115, s3 :: v_dual_and_b32 v96, 0xff, v96
+; GFX11-NEXT:    v_or_b32_e32 v6, v23, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_or_b32_e32 v11, v96, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 8, v10
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 8, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
+; GFX11-NEXT:    v_or_b32_e32 v23, v6, v11
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v21
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xff, v22
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v87
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xff, v26
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xff, v86
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
+; GFX11-NEXT:    v_or_b32_e32 v7, v11, v21
+; GFX11-NEXT:    v_or_b32_e32 v11, v22, v13
+; GFX11-NEXT:    v_or_b32_e32 v10, v26, v10
+; GFX11-NEXT:    v_or_b32_e32 v13, v24, v14
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v25
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v85
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xff, v29
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v84
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v27
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xff, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v83
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v21
+; GFX11-NEXT:    v_or_b32_e32 v15, v22, v15
+; GFX11-NEXT:    v_or_b32_e32 v9, v24, v9
+; GFX11-NEXT:    v_or_b32_e32 v16, v25, v16
+; GFX11-NEXT:    v_or_b32_e32 v21, v26, v27
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
+; GFX11-NEXT:    v_or_b32_e32 v24, v6, v7
+; GFX11-NEXT:    v_or_b32_e32 v25, v11, v10
+; GFX11-NEXT:    v_or_b32_e32 v26, v13, v14
+; GFX11-NEXT:    v_or_b32_e32 v6, v15, v9
+; GFX11-NEXT:    v_or_b32_e32 v7, v16, v21
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v32
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 8, v17
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xff, v82
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v31
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 8, v18
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v30
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v81
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v35
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v19
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v10
+; GFX11-NEXT:    v_or_b32_e32 v8, v11, v8
+; GFX11-NEXT:    v_or_b32_e32 v10, v13, v14
+; GFX11-NEXT:    v_or_b32_e32 v11, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v13, v17, v18
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v80
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v34
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v20
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v33
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v71
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v70
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
+; GFX11-NEXT:    v_or_b32_e32 v4, v14, v4
+; GFX11-NEXT:    v_or_b32_e32 v14, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v15, v17, v18
+; GFX11-NEXT:    v_or_b32_e32 v16, v19, v20
+; GFX11-NEXT:    v_or_b32_e32 v3, v21, v3
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_or_b32_e32 v8, v9, v8
+; GFX11-NEXT:    v_or_b32_e32 v9, v10, v11
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v4
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v15
+; GFX11-NEXT:    v_or_b32_e32 v15, v16, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 8, v68
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v67
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xff, v49
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v66
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xff, v65
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v39
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v64
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, v10, v11
+; GFX11-NEXT:    v_or_b32_e32 v10, v16, v17
+; GFX11-NEXT:    v_or_b32_e32 v2, v18, v2
+; GFX11-NEXT:    v_or_b32_e32 v11, v19, v20
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xff, v48
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v55
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xff, v52
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v54
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v53
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v51
+; GFX11-NEXT:    v_lshlrev_b32_e32 v12, 8, v12
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xff, v50
+; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
+; GFX11-NEXT:    v_or_b32_e32 v16, v16, v17
+; GFX11-NEXT:    v_or_b32_e32 v17, v18, v19
+; GFX11-NEXT:    v_or_b32_e32 v1, v20, v1
+; GFX11-NEXT:    v_or_b32_e32 v12, v21, v12
+; GFX11-NEXT:    v_or_b32_e32 v5, v22, v5
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT:    v_or_b32_e32 v16, v3, v4
+; GFX11-NEXT:    v_or_b32_e32 v1, v10, v2
+; GFX11-NEXT:    v_or_b32_e32 v2, v11, v18
+; GFX11-NEXT:    v_or_b32_e32 v3, v17, v19
+; GFX11-NEXT:    v_or_b32_e32 v4, v12, v5
+; GFX11-NEXT:    s_clause 0x5
+; GFX11-NEXT:    scratch_store_b128 v0, v[97:100], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[112:115], off offset:48
+; GFX11-NEXT:    scratch_store_b128 v0, v[23:26], off offset:64
+; GFX11-NEXT:    scratch_store_b128 v0, v[6:9], off offset:80
+; GFX11-NEXT:    scratch_store_b128 v0, v[13:16], off offset:96
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
+; GFX11-NEXT:    v_readlane_b32 s104, v41, 8
+; GFX11-NEXT:    v_readlane_b32 s103, v41, 7
+; GFX11-NEXT:    v_readlane_b32 s102, v41, 6
+; GFX11-NEXT:    v_readlane_b32 s101, v41, 5
+; GFX11-NEXT:    v_readlane_b32 s100, v41, 4
+; GFX11-NEXT:    v_readlane_b32 s99, v41, 3
+; GFX11-NEXT:    v_readlane_b32 s98, v41, 2
+; GFX11-NEXT:    v_readlane_b32 s97, v41, 1
+; GFX11-NEXT:    v_readlane_b32 s96, v41, 0
+; GFX11-NEXT:    v_readlane_b32 s87, v40, 31
+; GFX11-NEXT:    v_readlane_b32 s86, v40, 30
+; GFX11-NEXT:    v_readlane_b32 s85, v40, 29
+; GFX11-NEXT:    v_readlane_b32 s84, v40, 28
+; GFX11-NEXT:    v_readlane_b32 s83, v40, 27
+; GFX11-NEXT:    v_readlane_b32 s82, v40, 26
+; GFX11-NEXT:    v_readlane_b32 s81, v40, 25
+; GFX11-NEXT:    v_readlane_b32 s80, v40, 24
+; GFX11-NEXT:    v_readlane_b32 s71, v40, 23
+; GFX11-NEXT:    v_readlane_b32 s70, v40, 22
+; GFX11-NEXT:    v_readlane_b32 s69, v40, 21
+; GFX11-NEXT:    v_readlane_b32 s68, v40, 20
+; GFX11-NEXT:    v_readlane_b32 s67, v40, 19
+; GFX11-NEXT:    v_readlane_b32 s66, v40, 18
+; GFX11-NEXT:    v_readlane_b32 s65, v40, 17
+; GFX11-NEXT:    v_readlane_b32 s64, v40, 16
+; GFX11-NEXT:    v_readlane_b32 s55, v40, 15
+; GFX11-NEXT:    v_readlane_b32 s54, v40, 14
+; GFX11-NEXT:    v_readlane_b32 s53, v40, 13
+; GFX11-NEXT:    v_readlane_b32 s52, v40, 12
+; GFX11-NEXT:    v_readlane_b32 s51, v40, 11
+; GFX11-NEXT:    v_readlane_b32 s50, v40, 10
+; GFX11-NEXT:    v_readlane_b32 s49, v40, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v40, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v40, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32
+; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:12
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -188903,417 +179543,438 @@ define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64f16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:436
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:320
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v6 :: v_dual_mov_b32 v33, v0
+; GFX11-TRUE16-NEXT:    s_clause 0x1e
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:352
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:320
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, v30 :: v_dual_mov_b32 v51, v24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v28 :: v_dual_mov_b32 v55, v26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v22 :: v_dual_mov_b32 v48, v20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v18 :: v_dual_mov_b32 v39, v16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v14 :: v_dual_mov_b32 v34, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v12 :: v_dual_mov_b32 v38, v10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v6 :: v_dual_mov_b32 v33, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v2 :: v_dual_mov_b32 v32, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
 ; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:316
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v89, off, s32 offset:312
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v92, off, s32 offset:308
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v91, off, s32 offset:304
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v93, off, s32 offset:300
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:296
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:292
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v88, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v90, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v2, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v16, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v18, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v20, off, s32 offset:72
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v22, off, s32 offset:80
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v24, off, s32 offset:88
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v26, off, s32 offset:96
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v28, off, s32 offset:104
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v30, off, s32 offset:112
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:120
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:128
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:136
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:144
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:152
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:160
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:168
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:176
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:184
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:200
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:208
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:216
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:224
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:232
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:240
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:248
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:256
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v88, off, s32 offset:264
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v89, off, s32 offset:272
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v90, off, s32 offset:280
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v91, off, s32 offset:288
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v92, off, s32 offset:296
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v93, off, s32 offset:304
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v94, off, s32 offset:312
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:308
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:300
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:292
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:284
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:276
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:268
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:260
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:252
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:244
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:228
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:220
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:212
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:204
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:196
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:188
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:180
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:172
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:164
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:156
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:148
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:140
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:132
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:124
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:116
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:108
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:100
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:92
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:84
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:76
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:68
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:12
 ; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v86, v30 :: v_dual_mov_b32 v81, v29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v84, v28 :: v_dual_mov_b32 v83, v27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v85, v26 :: v_dual_mov_b32 v80, v24
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v69, v25 :: v_dual_mov_b32 v82, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v71, v23 :: v_dual_mov_b32 v64, v21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v67, v20 :: v_dual_mov_b32 v68, v19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v70, v18 :: v_dual_mov_b32 v55, v16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v17 :: v_dual_mov_b32 v65, v15
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v66, v14 :: v_dual_mov_b32 v49, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v12 :: v_dual_mov_b32 v54, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v11 :: v_dual_mov_b32 v36, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v8 :: v_dual_mov_b32 v37, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v7 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v32, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v101, 8, v6
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v102, 8, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v113, 8, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v112, 8, v12
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v130, 8, v14
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v128, 8, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v134, 8, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v132, 8, v20
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v161, 8, v22
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v147, 8, v24
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v166, 8, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v167, 8, v28
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v180, 8, v30
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v177, 8, v31
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v42, 8, v41
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(61)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 8, v44
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(60)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v45, 8, v45
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(59)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v44, 8, v56
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(58)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v59, 8, v59
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(57)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v56, 8, v60
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(56)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v60, 8, v61
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(55)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v61, 8, v62
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(54)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v63, 8, v63
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(53)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v62, 8, v72
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(52)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v73, 8, v73
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(51)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v72, 8, v74
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(50)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v75, 8, v75
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(49)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v74, 8, v76
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(48)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v77, 8, v77
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(47)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v76, 8, v78
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(46)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v78, 8, v79
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(45)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v79, 8, v88
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(44)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v89, 8, v89
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(43)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v88, 8, v90
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(42)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v91, 8, v91
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(41)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v90, 8, v92
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(40)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v92, 8, v93
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(39)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v93, 8, v94
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB93_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v33
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v32
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v35
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v2, 0xffff, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v1, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v53
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v6, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v49
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v66
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v65
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v55
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v68
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v3, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v2, 16, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v83
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v86
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v98
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v96
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v102
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v101
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v99
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v114
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v112
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v103
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v118
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v115
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v116
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v113
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v130
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v119
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v14, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v134
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v133
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v128
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v117
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v146
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v145
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v132
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v135
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v131
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v129
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v144
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v150
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v148
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v21
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v147
+; GFX11-TRUE16-NEXT:    v_and_b32_e64 v5, 0xffff, s5
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
 ; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v149
 ; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v151
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v160
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v161
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v163
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s26, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v164
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s27, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
 ; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v165
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v178
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v177
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v68
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v64
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v4, v67
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v37
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v70
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v7, v80
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v8, v81
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v9, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v86
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v83
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v96
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v85
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v10, v97
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v11, v87
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v99
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v3, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v103
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v114
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v98
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v0, 16, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v100
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v113
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v116
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v14, v128
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v117
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v102
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v130
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v133
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v14, v132
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v0, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v148
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v118
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v129
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v161
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v2, 16, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v166
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v144
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v134
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v147
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v167
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v17, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v22
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v151
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v149
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v20, 16, v21
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v167
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v177
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v165
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v162
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v42
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v41
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v178
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v115
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v45
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v44
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v131
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v119
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v59
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v56
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v176
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v182
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v181
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v179
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v135
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v60
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v61
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v180
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v42
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v40
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v183
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v150
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v146
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v63
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v62
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v41
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v43
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v44
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v163
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v160
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v73
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v72
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v45
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v47
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v56
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v176
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v164
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v75
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v74
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v57
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v62
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v61
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v59
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v179
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v77
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v76
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v60
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v74
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v73
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v63
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v183
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v182
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v78
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v79
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v72
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v77
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v76
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v75
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v43
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v40
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v89
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v88
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v78
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v90
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v88
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v79
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v47
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v46
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v91
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v90
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v89
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v93
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v91
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v92
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v58
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v57
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v92
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v93
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v1, 16, v0
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
@@ -189360,396 +180021,344 @@ define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v92
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v93
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(37)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v58
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v57
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(35)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v47
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v90
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v79
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v92, v0
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v91
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v88
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v78
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v77
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v7, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v46
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v93, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v91, v2
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(33)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v43
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v5, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v76
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v75
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v74
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v72
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v73
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v40
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v90, v3
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(31)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v183
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v182
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v63
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(59)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v62
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v61
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v59
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v59, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v60, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v57
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(55)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v58
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v57, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v56
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v56, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v47
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v89, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(29)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v181
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v88, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v181, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v78, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v79, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v179
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v182, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v179, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v77, v3
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(27)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v176
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v45
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(51)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v46
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v45, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v44
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v43
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v41
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(47)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v42
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v164
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(25)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v163
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v163, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v76, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v160
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v75, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v160, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v74, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v73, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v1
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(23)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v150
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v40
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v183
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(43)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v182
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v180
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v181
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(39)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v178
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v176
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v72, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v146
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v177
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v167
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v167, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v176, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v165
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(35)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v166
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v165, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v164
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v164, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v163
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(21)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v145
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v135
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v63, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(19)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v131
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v62, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v131, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v60, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v61, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v119
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v135, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v119, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v59, v3
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(17)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v178
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v161
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v162
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v161, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v160
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v151
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v149
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v150
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v115
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v165
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v115, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v56, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v162
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v45, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v145, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v44, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v42, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v1
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v151
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v148
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v147
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(23)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v146
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v144
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v145
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v135
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(19)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v134
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v132
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v41, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v149
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v133
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v131
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v131, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v132, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v129
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v130
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v129, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v128
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v128, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v119
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v117
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(11)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v118
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v117, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v116
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v115
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v113
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v148
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v144
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v180, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(9)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v133
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v177, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v166, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v167, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v133, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v129
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v129, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v144, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v161, v3
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v114
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v112
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v103
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v102
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v101
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v99
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v86
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v118
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v117
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(5)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v116
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v116, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v147, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v114
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v96
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v55, 3, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v99
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v114, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v132, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v130, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v103
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v98
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v54
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v39
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 3, v33
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v113, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v128, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v100
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v101, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v102, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v98
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v96
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v134, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v97, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v87
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v85
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v84
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v83
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v81
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v82
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v87, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v51
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v86, v4
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v69
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v70
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v70, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v69, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v68
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v67
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v66
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v64
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v85, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v84, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v49
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v83, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v48
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v65
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v64, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v52
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v82, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v38
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v81, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v71, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v80, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v37
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v70, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v36
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, 3, v51
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v53, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v52, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v49
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v50
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 3, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v38, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v36, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v37, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v33
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v35
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v69, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v112, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v68, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v67, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v66, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v33
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v34, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v32
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v65, v8
 ; GFX11-TRUE16-NEXT:    v_and_b32_e64 v8, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v64
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v64, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v7, 16, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v32, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v51
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v38
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v53, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v49, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v69
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v33, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v34, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v50, 16, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v132
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v3, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v116
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v129
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v18
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v114, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v144, 16, v33
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v128
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v129, 16, v33
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v176
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v117, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v164
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v165, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v161, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v56
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v14, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v115
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v135
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v131
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v23
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v27
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v145, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v119, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v163
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v182
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v28
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v2, 16, v0
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v131, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v167, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v45, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v57, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v59, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v133, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v160, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v179, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v36
 ; GFX11-TRUE16-NEXT:  .LBB93_3: ; %end
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:320
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    s_clause 0x1e
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:320
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:352
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:440
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ; GFX11-TRUE16-NEXT:  .LBB93_4:
@@ -200422,2095 +191031,1034 @@ define inreg <128 x i8> @bitcast_v64f16_to_v128i8_scalar(<64 x half> inreg %a, i
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v64f16_to_v128i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:88
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s30, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s96, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s97, 1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s98, 2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s99, 3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s100, 4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s101, 5
-; GFX11-TRUE16-NEXT:    s_mov_b32 vcc_hi, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_clause 0x12
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s102, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s103, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s104, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s50, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s51, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s52, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s53, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s54, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s55, 15
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s64, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s65, 17
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s66, 18
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s67, 19
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s68, 20
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s69, 21
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s70, 22
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s71, 23
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s80, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s81, 25
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s82, 26
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s83, 27
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s84, 28
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s85, 29
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s86, 30
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s87, 31
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB95_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[0:1], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s5, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s5, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s5, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s4, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s7, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s7, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s51, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s52, s6, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s53, s9, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s54, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s55, s9, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s64, s8, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s65, s8, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s66, s11, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s67, s11, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s68, s11, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s69, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s70, s10, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s71, s13, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s80, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s81, s13, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s82, s12, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 31
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s83, s12, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s84, s15, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s85, s15, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 30
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s86, s15, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s87, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s96, s14, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 29
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s97, s41, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s98, s41, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s41, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 28
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s100, s40, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s101, s40, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s29, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 27
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s29, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s28, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 26
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[62:63], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[72:73], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 25
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 23
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[88:89], s[10:11], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[14:15], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 22
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[40:41], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 21
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 20
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 19
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s18, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 18
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s18, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 17
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 15
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 14
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 13
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s16, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 11
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s2, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s2, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s0, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s0, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s74, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s75, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[4:5], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB95_4
-; GFX11-TRUE16-NEXT:  .LBB95_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v51, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v50, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v36, 0x200, s19 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s21 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v28, 0x200, s27 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v27, 0x200, s26 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v35, 0x200, s18 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v22, 0x200, s29 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, s28 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, s41 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s40 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s4 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v55, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v54, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v49, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v48, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s23 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v32, 0x200, s25 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v31, 0x200, s24 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s22 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[23:24], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[27:28]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[35:36]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[80:81], 24, v[50:51]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[24:25], 24, v[9:10]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[31:32]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[48:49]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[81:82], 24, v[54:55]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[5:6]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[25:26], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[29:30], 24, v[13:14]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[17:18]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[21:22]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 24, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v164, 8, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v165, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v166, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v167, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v176, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v177, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v178, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v179, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v180, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v183, 24, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v40, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v41, 8, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v44, 24, v49
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v45, 16, v49
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v46, 8, v49
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v57, 24, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v58, 16, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v60, 16, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v61, 8, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v62, 24, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v63, 16, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v72, 8, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v54
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v54
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 24, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 24, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 24, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 24, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 8, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v21
-; GFX11-TRUE16-NEXT:    s_branch .LBB95_5
-; GFX11-TRUE16-NEXT:  .LBB95_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr104_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr103_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr102_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr101_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr99_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr98_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr97_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr96_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr87_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr86_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr85_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr84_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s43, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 8
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 9
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 10
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 11
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 12
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 13
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 14
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 15
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 17
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 18
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 19
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 20
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 21
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 22
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 23
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 24
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 25
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 26
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 27
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 28
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 29
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 30
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 2
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 8
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB95_2
-; GFX11-TRUE16-NEXT:  .LBB95_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v74.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v73.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v72.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v63.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v61.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v150.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v59.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v149.l, s34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v144.l, s104
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v135.l, s103
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v134.l, s102
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v133.l, s101
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v132.l, s100
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v131.l, s99
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v130.l, s98
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v129.l, s97
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.l, s96
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v119.l, s87
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v118.l, s86
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v56.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v117.l, s85
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v116.l, s84
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v115.l, s83
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v114.l, s82
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v113.l, s81
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v112.l, s80
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v103.l, s71
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v102.l, s70
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v46.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v101.l, s69
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v100.l, s68
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v99.l, s67
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v98.l, s66
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v97.l, s65
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v96.l, s64
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, s55
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v86.l, s54
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v44.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.l, s53
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.l, s52
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v83.l, s51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v82.l, s50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v43.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v71.l, s49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v145.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v41.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s72
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v70.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v80.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s74
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v183.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s76
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s78
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s88
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s90
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v182.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s92
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, s94
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v68.l, s30
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v181.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v180.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 25
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v179.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v178.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 27
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v177.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v176.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 29
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v167.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 30
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v166.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 31
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v165.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v164.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v163.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v162.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v161.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v160.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v151.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v148.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v147.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v146.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s1, v78, 3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v81.l, s0
-; GFX11-TRUE16-NEXT:  .LBB95_5: ; %end
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v74
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xff, v54
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xff, v55
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v62
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, v54, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v73
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xff, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v60, 0xff, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v69, v66
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v72
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v80, v60, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v57, 8, v57
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, v55, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v63
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xff, v49
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, v54, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, v69, v81
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v61
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v50, v81
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v59
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v81, 0xff, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v64
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, v55, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v51, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, v81, v57
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v56
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v81, 0xff, v47
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v56, v50, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, v48, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v81, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v46
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xff, v45
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v44
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v57, v50, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, v49, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v70, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v43
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v42
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, v48, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, v35, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v69, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v41
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v40
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v183
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, v49, v50
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v36, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v67, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v181
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v182
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v64, v67, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v35, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v36, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xff, v179
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v178
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v177
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v176
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v167
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v36, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v52, v67, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[54:57], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[48:51], off offset:16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v164
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xff, v163
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v162
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xff, v161
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v160
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v151
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v66, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v31, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, v36, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v32, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v49, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v27, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xff, v148
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v147
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xff, v146
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v145
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v150
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xff, v149
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, v38, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v28, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v49, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, v52, v54
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, v31, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, v27, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v28, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, v21, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v144
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v135
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v134
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v133
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v32, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v132
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 8, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v131
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v27, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v17, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v130
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v129
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v32, v49
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v128
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xff, v119
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v28, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v118
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xff, v117
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v116
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v49, v33
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v51, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, v21, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v17, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v18, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v115
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v114
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v113
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v102
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v112
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v103
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v18, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v101
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v96
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v27, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v99
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v98
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v31, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v27, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v14, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v9, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v86
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v85
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xff, v166
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v165
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v84
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v14, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v24, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v64, v70, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v53
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v24, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v26, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v67, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v23, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v24, v20
-; GFX11-TRUE16-NEXT:    s_clause 0x5
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[35:38], off offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[48:51], off offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[15:18], off offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[11:14], off offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:112
-; GFX11-TRUE16-NEXT:    s_clause 0x12
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s104, v76, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s103, v76, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s102, v76, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s101, v76, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s100, v76, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s99, v76, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s98, v76, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s97, v76, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s96, v76, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s87, v75, 31
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s86, v75, 30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s85, v75, 29
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s84, v75, 28
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s83, v75, 27
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s82, v75, 26
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s81, v75, 25
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s80, v75, 24
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s71, v75, 23
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s70, v75, 22
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s69, v75, 21
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s68, v75, 20
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v75, 19
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v75, 18
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v75, 17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v75, 16
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v75, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v75, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v75, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v75, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v75, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v75, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v75, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v75, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v75, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v75, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v75, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v75, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v75, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v75, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v75, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v75, 0
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v64f16_to_v128i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:88
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s96, 0
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s97, 1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s98, 2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s99, 3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s100, 4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s101, 5
-; GFX11-FAKE16-NEXT:    s_mov_b32 s99, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_clause 0x12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s102, 6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s103, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s104, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s49, 9
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s50, 10
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s51, 11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s52, 12
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s53, 13
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s54, 14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s55, 15
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s64, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s65, 17
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s66, 18
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s67, 19
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s68, 20
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s69, 21
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s70, 22
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s71, 23
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s80, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s81, 25
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s82, 26
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s83, 27
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s84, 28
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s85, 29
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s86, 30
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s87, 31
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB95_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[74:75], s[0:1], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s5, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s5, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s5, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s4, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s7, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s50, s7, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s51, s6, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s52, s9, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s53, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s55, s9, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s54, s8, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s64, s8, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s65, s11, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s66, s11, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s68, s11, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s67, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s69, s10, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s70, s13, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s71, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s81, s13, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s80, s12, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 31
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s82, s12, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s83, s15, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s84, s15, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 30
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s86, s15, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s85, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s87, s14, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 29
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s96, s41, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s97, s41, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s41, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 28
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s98, s40, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s101, s40, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s29, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 27
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s29, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s28, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 26
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[72:73], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 25
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 23
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[88:89], s[10:11], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[92:93], s[14:15], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 22
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[40:41], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 21
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 20
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 19
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 18
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 17
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 15
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 14
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 13
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 11
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 10
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s2, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s2, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s0, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s0, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s74, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s75, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[74:75], s[4:5], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s99
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB95_4
-; GFX11-FAKE16-NEXT:  .LBB95_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v39, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v38, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v1, 0x200, s4 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v51, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v50, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v33, 0x200, s21 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v29, 0x200, s23 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v28, 0x200, s22 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v32, 0x200, s20 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v16, 0x200, s29 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v15, 0x200, s28 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v14, 0x200, s41 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v13, 0x200, s40 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v12, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v11, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v10, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v9, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v8, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v7, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v6, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v5, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v4, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v3, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v53, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v52, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v37, 0x200, s19 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v25, 0x200, s25 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v21, 0x200, s27 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v20, 0x200, s26 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v24, 0x200, s24 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v36, 0x200, s18 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[80:81], 24, v[38:39]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[28:29]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[32:33]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[81:82], 24, v[50:51]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[20:21]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[24:25]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[36:37]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[82:83], 24, v[52:53]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[18:19], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[7:8]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[30:31], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[11:12]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[15:16]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 8, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 8, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 8, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 8, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 16, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 8, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 8, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v51
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v51
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v51
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v50
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v50
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 16, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 8, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v52
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v52
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 24, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 24, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 24, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v15
-; GFX11-FAKE16-NEXT:    s_branch .LBB95_5
-; GFX11-FAKE16-NEXT:  .LBB95_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    s_mov_b32 s99, -1
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr104
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr103
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr102
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr101
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr98
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr100
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr97
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr96
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr87
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr85
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr86
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s43, 1
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    s_branch .LBB95_2
-; GFX11-FAKE16-NEXT:  .LBB95_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s0 :: v_dual_mov_b32 v53, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v71, s50
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v74, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s12 :: v_dual_mov_b32 v10, s13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v73, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v55, s48
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v72, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v49, s39
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s6 :: v_dual_mov_b32 v4, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s4 :: v_dual_mov_b32 v2, s5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v62, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, s2 :: v_dual_mov_b32 v51, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s16 :: v_dual_mov_b32 v39, s17
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v63, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v35, s38 :: v_dual_mov_b32 v36, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v37, s19 :: v_dual_mov_b32 v32, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v33, s21 :: v_dual_mov_b32 v60, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s22 :: v_dual_mov_b32 v29, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v61, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s26 :: v_dual_mov_b32 v21, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v146, s42 :: v_dual_mov_b32 v145, s104
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v59, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v144, vcc_hi :: v_dual_mov_b32 v135, s103
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v134, s102 :: v_dual_mov_b32 v133, s101
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v57, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v31, s36 :: v_dual_mov_b32 v132, s98
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v131, s100 :: v_dual_mov_b32 v130, s97
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v129, s96 :: v_dual_mov_b32 v58, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s37 :: v_dual_mov_b32 v128, s87
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v119, s85 :: v_dual_mov_b32 v118, s86
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v117, s84 :: v_dual_mov_b32 v56, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v116, s83 :: v_dual_mov_b32 v115, s82
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v114, s80 :: v_dual_mov_b32 v113, s81
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v47, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s35 :: v_dual_mov_b32 v112, s71
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v103, s70 :: v_dual_mov_b32 v102, s69
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v101, s67 :: v_dual_mov_b32 v46, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s34 :: v_dual_mov_b32 v100, s68
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v99, s66 :: v_dual_mov_b32 v98, s65
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v97, s64 :: v_dual_mov_b32 v44, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v96, s54 :: v_dual_mov_b32 v87, s55
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s53 :: v_dual_mov_b32 v85, s52
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v45, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v84, s51 :: v_dual_mov_b32 v83, s49
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v147, s43 :: v_dual_mov_b32 v22, s78
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v43, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v67, s58 :: v_dual_mov_b32 v26, s88
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v81, s44 :: v_dual_mov_b32 v30, s90
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v42, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s74 :: v_dual_mov_b32 v34, s92
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v65, s94 :: v_dual_mov_b32 v68, s30
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v41, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 20
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v48, s62
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v54, s72
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v64, s60
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, s56 :: v_dual_mov_b32 v183, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 21
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v80, s46
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v18, s76
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v40, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 22
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v182, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 23
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v181, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v180, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 25
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v178, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 26
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v179, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v177, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 28
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v176, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 29
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v167, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 30
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v165, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 31
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v166, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v164, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v163, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v162, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v160, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v161, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v151, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v150, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 7
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v149, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v148, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s1, v78, 1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v82, s0
-; GFX11-FAKE16-NEXT:  .LBB95_5: ; %end
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v74
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xff, v52
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xff, v53
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v63
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v52, v69
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xff, v73
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xff, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v57, 0xff, v57
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v58, 8, v58
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v69, v66
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v72
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v80
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v53, v69
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xff, v62
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v67
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v69, v82
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v60
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v60, 0xff, v61
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v64
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, v50, v82
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v81, v60, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, v52, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v53
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v69
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xff, v51
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v59
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, v52, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v82
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v66, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v57, v58
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v56
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xff, v47
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v52, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xffff, v66
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, v38, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v82, v80
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v46
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v81, 0xff, v44
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v45
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v53, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, v39, v80
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v81, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v80, 0xff, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v43
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xff, v42
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v38, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v39
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v80, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v82, v70
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v70, 0xff, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v41
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v81, 0xff, v183
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v40
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v38, v39
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v66
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v70, v80
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v81, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v80, 0xff, v181
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v182
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v180
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xffff, v66
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, v80, v67
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v32, v70
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v33, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, v38, v39
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, v66, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v67
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v67, 0xff, v178
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v179
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v177
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v80, 0xff, v176
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v167
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xff, v165
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v165, 8, v166
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, v67, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v70
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, v80, v64
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v29, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v82, v165
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v67
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[50:53], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[36:39], off offset:16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v32, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v33, v67
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, v28, v64
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, v29, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v164
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v163
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v54
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v162
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xff, v160
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v161
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v151
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v25, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v50, v51
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v52
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v150
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v149
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xff, v148
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v147
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v146
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xff, v145
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v68
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v32, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v50, v51
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, v53, v54
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, v24, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v20, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v21, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, v15, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v144
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v135
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v134
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v133
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, v25, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v132
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v65
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v131
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v130
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v129
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v25, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v128
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v119
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v118
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v117
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v116
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v28, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v33, v34
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v13, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, v14, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v115
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v114
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v30
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v113
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v112
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v103
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v24
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v102
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v20, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v101
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v100
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v97
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v96
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v99
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v98
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v26, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v21, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v9, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v10, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v5, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v87
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v86
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v85
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v83
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v84
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v71
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v10, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v20, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v49
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v55
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v31
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v23
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v22, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v19, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v20, v17
-; GFX11-FAKE16-NEXT:    s_clause 0x5
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[36:39], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[50:53], off offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[64:67], off offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
-; GFX11-FAKE16-NEXT:    s_clause 0x12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:72
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s104, v76, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s103, v76, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s102, v76, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s101, v76, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s100, v76, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s99, v76, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s98, v76, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s97, v76, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s96, v76, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s87, v75, 31
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s86, v75, 30
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s85, v75, 29
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s84, v75, 28
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s83, v75, 27
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s82, v75, 26
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s81, v75, 25
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s80, v75, 24
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s71, v75, 23
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s70, v75, 22
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s69, v75, 21
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s68, v75, 20
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v75, 19
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v75, 18
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v75, 17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v75, 16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v75, 15
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v75, 14
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v75, 13
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v75, 12
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v75, 11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v75, 10
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v75, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v75, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v75, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v75, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v75, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v75, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v75, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v75, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v75, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v75, 0
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:88
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v64f16_to_v128i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:76
+; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:80
+; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:84
+; GFX11-NEXT:    scratch_store_b32 off, v78, s32 offset:88
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v75, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v76, s96, 0
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
+; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s41, v2
+; GFX11-NEXT:    v_writelane_b32 v75, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v76, s97, 1
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s15, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
+; GFX11-NEXT:    v_writelane_b32 v75, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v76, s98, 2
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v8
+; GFX11-NEXT:    v_writelane_b32 v75, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v76, s99, 3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
+; GFX11-NEXT:    v_writelane_b32 v75, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v76, s100, 4
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v14
+; GFX11-NEXT:    v_writelane_b32 v75, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v76, s101, 5
+; GFX11-NEXT:    s_mov_b32 s99, 0
+; GFX11-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
+; GFX11-NEXT:    s_clause 0x12
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:72
+; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:68
+; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:64
+; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:60
+; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:56
+; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:52
+; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:48
+; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:44
+; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:40
+; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:36
+; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:32
+; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:28
+; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:24
+; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:20
+; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:16
+; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:12
+; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v73, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v74, s32
+; GFX11-NEXT:    v_writelane_b32 v75, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v76, s102, 6
+; GFX11-NEXT:    ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
+; GFX11-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
+; GFX11-NEXT:    v_writelane_b32 v75, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v76, s103, 7
+; GFX11-NEXT:    v_writelane_b32 v75, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v76, s104, 8
+; GFX11-NEXT:    v_writelane_b32 v75, s49, 9
+; GFX11-NEXT:    v_writelane_b32 v75, s50, 10
+; GFX11-NEXT:    v_writelane_b32 v75, s51, 11
+; GFX11-NEXT:    v_writelane_b32 v75, s52, 12
+; GFX11-NEXT:    v_writelane_b32 v75, s53, 13
+; GFX11-NEXT:    v_writelane_b32 v75, s54, 14
+; GFX11-NEXT:    v_writelane_b32 v75, s55, 15
+; GFX11-NEXT:    v_writelane_b32 v75, s64, 16
+; GFX11-NEXT:    v_writelane_b32 v75, s65, 17
+; GFX11-NEXT:    v_writelane_b32 v75, s66, 18
+; GFX11-NEXT:    v_writelane_b32 v75, s67, 19
+; GFX11-NEXT:    v_writelane_b32 v75, s68, 20
+; GFX11-NEXT:    v_writelane_b32 v75, s69, 21
+; GFX11-NEXT:    v_writelane_b32 v75, s70, 22
+; GFX11-NEXT:    v_writelane_b32 v75, s71, 23
+; GFX11-NEXT:    v_writelane_b32 v75, s80, 24
+; GFX11-NEXT:    v_writelane_b32 v75, s81, 25
+; GFX11-NEXT:    v_writelane_b32 v75, s82, 26
+; GFX11-NEXT:    v_writelane_b32 v75, s83, 27
+; GFX11-NEXT:    v_writelane_b32 v75, s84, 28
+; GFX11-NEXT:    v_writelane_b32 v75, s85, 29
+; GFX11-NEXT:    v_writelane_b32 v75, s86, 30
+; GFX11-NEXT:    v_writelane_b32 v75, s87, 31
+; GFX11-NEXT:    s_cbranch_scc0 .LBB95_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b64 s[74:75], s[0:1], 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s34, s5, 24
+; GFX11-NEXT:    s_lshr_b32 s35, s5, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 7
+; GFX11-NEXT:    s_lshr_b32 s42, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s37, s5, 8
+; GFX11-NEXT:    s_lshr_b32 s36, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s38, s4, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 6
+; GFX11-NEXT:    s_lshr_b32 s42, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s7, 24
+; GFX11-NEXT:    s_lshr_b32 s48, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s50, s7, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 5
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s49, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s51, s6, 8
+; GFX11-NEXT:    s_lshr_b32 s52, s9, 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 4
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s53, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s55, s9, 8
+; GFX11-NEXT:    s_lshr_b32 s54, s8, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s64, s8, 8
+; GFX11-NEXT:    s_lshr_b32 s65, s11, 24
+; GFX11-NEXT:    s_lshr_b32 s66, s11, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 2
+; GFX11-NEXT:    s_lshr_b32 s42, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s68, s11, 8
+; GFX11-NEXT:    s_lshr_b32 s67, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s69, s10, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 1
+; GFX11-NEXT:    s_lshr_b32 s42, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s70, s13, 24
+; GFX11-NEXT:    s_lshr_b32 s71, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s81, s13, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 0
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s80, s12, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 31
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s82, s12, 8
+; GFX11-NEXT:    s_lshr_b32 s83, s15, 24
+; GFX11-NEXT:    s_lshr_b32 s84, s15, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 30
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s86, s15, 8
+; GFX11-NEXT:    s_lshr_b32 s85, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s87, s14, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 29
+; GFX11-NEXT:    s_lshr_b32 s42, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s96, s41, 24
+; GFX11-NEXT:    s_lshr_b32 s97, s41, 16
+; GFX11-NEXT:    s_lshr_b32 s100, s41, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 28
+; GFX11-NEXT:    s_lshr_b32 s42, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s98, s40, 16
+; GFX11-NEXT:    s_lshr_b32 s101, s40, 8
+; GFX11-NEXT:    s_lshr_b32 s102, s29, 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 27
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s103, s29, 16
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s29, 8
+; GFX11-NEXT:    s_lshr_b32 s104, s28, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 26
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 16
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[72:73], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 25
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 8
+; GFX11-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s20, 16
+; GFX11-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
+; GFX11-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 23
+; GFX11-NEXT:    s_lshr_b32 s42, s20, 8
+; GFX11-NEXT:    s_lshr_b64 s[88:89], s[10:11], 24
+; GFX11-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
+; GFX11-NEXT:    s_lshr_b64 s[92:93], s[14:15], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 22
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 24
+; GFX11-NEXT:    s_lshr_b64 s[94:95], s[40:41], 24
+; GFX11-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 21
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 20
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 19
+; GFX11-NEXT:    s_lshr_b32 s42, s18, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 18
+; GFX11-NEXT:    s_lshr_b32 s42, s18, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 17
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 15
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 14
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 13
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 12
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 11
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 10
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 9
+; GFX11-NEXT:    s_lshr_b32 s42, s2, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s2, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 7
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 6
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 5
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 4
+; GFX11-NEXT:    s_lshr_b32 s42, s0, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s0, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 2
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s74, 0
+; GFX11-NEXT:    v_writelane_b32 v78, s75, 1
+; GFX11-NEXT:    s_lshr_b64 s[74:75], s[4:5], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s99
+; GFX11-NEXT:    s_cbranch_vccnz .LBB95_4
+; GFX11-NEXT:  .LBB95_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v39, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v38, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v51, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v50, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v33, 0x200, s21 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, s23 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, s22 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v32, 0x200, s20 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s29 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s28 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s41 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s40 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v53, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v52, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v37, 0x200, s19 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, s25 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, s27 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, s26 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, s24 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v36, 0x200, s18 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_lshrrev_b64 v[80:81], 24, v[38:39]
+; GFX11-NEXT:    v_lshrrev_b64 v[64:65], 24, v[28:29]
+; GFX11-NEXT:    v_lshrrev_b64 v[67:68], 24, v[32:33]
+; GFX11-NEXT:    v_lshrrev_b64 v[81:82], 24, v[50:51]
+; GFX11-NEXT:    v_lshrrev_b64 v[17:18], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[48:49], 24, v[20:21]
+; GFX11-NEXT:    v_lshrrev_b64 v[54:55], 24, v[24:25]
+; GFX11-NEXT:    v_lshrrev_b64 v[70:71], 24, v[36:37]
+; GFX11-NEXT:    v_lshrrev_b64 v[82:83], 24, v[52:53]
+; GFX11-NEXT:    v_lshrrev_b64 v[18:19], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[22:23], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b64 v[26:27], 24, v[7:8]
+; GFX11-NEXT:    v_lshrrev_b64 v[30:31], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[34:35], 24, v[11:12]
+; GFX11-NEXT:    v_lshrrev_b64 v[65:66], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b64 v[68:69], 24, v[15:16]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v147, 24, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v148, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v149, 8, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v150, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v151, 8, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v161, 24, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v160, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v162, 8, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v163, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v164, 8, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v166, 24, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v165, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v167, 8, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v176, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v177, 8, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v179, 24, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v178, 16, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v180, 8, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v181, 16, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v182, 8, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v40, 24, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v183, 16, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v41, 8, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v42, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v43, 8, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v45, 24, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v44, 16, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v46, 8, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v47, 16, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v56, 8, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v58, 24, v51
+; GFX11-NEXT:    v_lshrrev_b32_e32 v57, 16, v51
+; GFX11-NEXT:    v_lshrrev_b32_e32 v59, 8, v51
+; GFX11-NEXT:    v_lshrrev_b32_e32 v61, 16, v50
+; GFX11-NEXT:    v_lshrrev_b32_e32 v60, 8, v50
+; GFX11-NEXT:    v_lshrrev_b32_e32 v63, 24, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v62, 16, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v72, 8, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v73, 16, v52
+; GFX11-NEXT:    v_lshrrev_b32_e32 v74, 8, v52
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 24, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 24, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 8, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v97, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v98, 24, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v99, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v100, 8, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v101, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v102, 8, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v103, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v112, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v113, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v114, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v115, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v116, 24, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v117, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v118, 8, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v119, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v128, 8, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v129, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v130, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v131, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v132, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v133, 8, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v134, 24, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v135, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v144, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v145, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v146, 8, v15
+; GFX11-NEXT:    s_branch .LBB95_5
+; GFX11-NEXT:  .LBB95_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    s_mov_b32 s99, -1
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr104
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $sgpr103
+; GFX11-NEXT:    ; implicit-def: $sgpr102
+; GFX11-NEXT:    ; implicit-def: $sgpr101
+; GFX11-NEXT:    ; implicit-def: $sgpr98
+; GFX11-NEXT:    ; implicit-def: $sgpr100
+; GFX11-NEXT:    ; implicit-def: $sgpr97
+; GFX11-NEXT:    ; implicit-def: $sgpr96
+; GFX11-NEXT:    ; implicit-def: $sgpr87
+; GFX11-NEXT:    ; implicit-def: $sgpr85
+; GFX11-NEXT:    ; implicit-def: $sgpr86
+; GFX11-NEXT:    ; implicit-def: $sgpr84
+; GFX11-NEXT:    ; implicit-def: $sgpr83
+; GFX11-NEXT:    ; implicit-def: $sgpr82
+; GFX11-NEXT:    ; implicit-def: $sgpr80
+; GFX11-NEXT:    ; implicit-def: $sgpr81
+; GFX11-NEXT:    ; implicit-def: $sgpr71
+; GFX11-NEXT:    ; implicit-def: $sgpr70
+; GFX11-NEXT:    ; implicit-def: $sgpr69
+; GFX11-NEXT:    ; implicit-def: $sgpr67
+; GFX11-NEXT:    ; implicit-def: $sgpr68
+; GFX11-NEXT:    ; implicit-def: $sgpr66
+; GFX11-NEXT:    ; implicit-def: $sgpr65
+; GFX11-NEXT:    ; implicit-def: $sgpr64
+; GFX11-NEXT:    ; implicit-def: $sgpr54
+; GFX11-NEXT:    ; implicit-def: $sgpr55
+; GFX11-NEXT:    ; implicit-def: $sgpr53
+; GFX11-NEXT:    ; implicit-def: $sgpr52
+; GFX11-NEXT:    ; implicit-def: $sgpr51
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr50
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 0
+; GFX11-NEXT:    v_writelane_b32 v78, s43, 1
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    s_branch .LBB95_2
+; GFX11-NEXT:  .LBB95_4:
+; GFX11-NEXT:    v_dual_mov_b32 v52, s0 :: v_dual_mov_b32 v53, s1
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 2
+; GFX11-NEXT:    v_mov_b32_e32 v71, s50
+; GFX11-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v74, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 3
+; GFX11-NEXT:    v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
+; GFX11-NEXT:    v_dual_mov_b32 v9, s12 :: v_dual_mov_b32 v10, s13
+; GFX11-NEXT:    v_mov_b32_e32 v73, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 4
+; GFX11-NEXT:    v_mov_b32_e32 v55, s48
+; GFX11-NEXT:    v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
+; GFX11-NEXT:    v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_mov_b32_e32 v72, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 5
+; GFX11-NEXT:    v_mov_b32_e32 v49, s39
+; GFX11-NEXT:    v_dual_mov_b32 v3, s6 :: v_dual_mov_b32 v4, s7
+; GFX11-NEXT:    v_dual_mov_b32 v1, s4 :: v_dual_mov_b32 v2, s5
+; GFX11-NEXT:    v_mov_b32_e32 v62, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 6
+; GFX11-NEXT:    v_dual_mov_b32 v50, s2 :: v_dual_mov_b32 v51, s3
+; GFX11-NEXT:    v_dual_mov_b32 v38, s16 :: v_dual_mov_b32 v39, s17
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v63, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 7
+; GFX11-NEXT:    v_dual_mov_b32 v35, s38 :: v_dual_mov_b32 v36, s18
+; GFX11-NEXT:    v_dual_mov_b32 v37, s19 :: v_dual_mov_b32 v32, s20
+; GFX11-NEXT:    v_dual_mov_b32 v33, s21 :: v_dual_mov_b32 v60, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 8
+; GFX11-NEXT:    v_dual_mov_b32 v28, s22 :: v_dual_mov_b32 v29, s23
+; GFX11-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v61, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 9
+; GFX11-NEXT:    v_dual_mov_b32 v20, s26 :: v_dual_mov_b32 v21, s27
+; GFX11-NEXT:    v_dual_mov_b32 v146, s42 :: v_dual_mov_b32 v145, s104
+; GFX11-NEXT:    v_mov_b32_e32 v59, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 10
+; GFX11-NEXT:    v_dual_mov_b32 v144, vcc_hi :: v_dual_mov_b32 v135, s103
+; GFX11-NEXT:    v_dual_mov_b32 v134, s102 :: v_dual_mov_b32 v133, s101
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v57, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 11
+; GFX11-NEXT:    v_dual_mov_b32 v31, s36 :: v_dual_mov_b32 v132, s98
+; GFX11-NEXT:    v_dual_mov_b32 v131, s100 :: v_dual_mov_b32 v130, s97
+; GFX11-NEXT:    v_dual_mov_b32 v129, s96 :: v_dual_mov_b32 v58, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 12
+; GFX11-NEXT:    v_dual_mov_b32 v27, s37 :: v_dual_mov_b32 v128, s87
+; GFX11-NEXT:    v_dual_mov_b32 v119, s85 :: v_dual_mov_b32 v118, s86
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v117, s84 :: v_dual_mov_b32 v56, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 13
+; GFX11-NEXT:    v_dual_mov_b32 v116, s83 :: v_dual_mov_b32 v115, s82
+; GFX11-NEXT:    v_dual_mov_b32 v114, s80 :: v_dual_mov_b32 v113, s81
+; GFX11-NEXT:    v_mov_b32_e32 v47, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 14
+; GFX11-NEXT:    v_dual_mov_b32 v23, s35 :: v_dual_mov_b32 v112, s71
+; GFX11-NEXT:    v_dual_mov_b32 v103, s70 :: v_dual_mov_b32 v102, s69
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v101, s67 :: v_dual_mov_b32 v46, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 15
+; GFX11-NEXT:    v_dual_mov_b32 v19, s34 :: v_dual_mov_b32 v100, s68
+; GFX11-NEXT:    v_dual_mov_b32 v99, s66 :: v_dual_mov_b32 v98, s65
+; GFX11-NEXT:    v_dual_mov_b32 v97, s64 :: v_dual_mov_b32 v44, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 16
+; GFX11-NEXT:    v_dual_mov_b32 v96, s54 :: v_dual_mov_b32 v87, s55
+; GFX11-NEXT:    v_dual_mov_b32 v86, s53 :: v_dual_mov_b32 v85, s52
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v45, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 17
+; GFX11-NEXT:    v_dual_mov_b32 v84, s51 :: v_dual_mov_b32 v83, s49
+; GFX11-NEXT:    v_dual_mov_b32 v147, s43 :: v_dual_mov_b32 v22, s78
+; GFX11-NEXT:    v_mov_b32_e32 v43, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 18
+; GFX11-NEXT:    v_dual_mov_b32 v67, s58 :: v_dual_mov_b32 v26, s88
+; GFX11-NEXT:    v_dual_mov_b32 v81, s44 :: v_dual_mov_b32 v30, s90
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v42, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 19
+; GFX11-NEXT:    v_dual_mov_b32 v17, s74 :: v_dual_mov_b32 v34, s92
+; GFX11-NEXT:    v_dual_mov_b32 v65, s94 :: v_dual_mov_b32 v68, s30
+; GFX11-NEXT:    v_mov_b32_e32 v41, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 20
+; GFX11-NEXT:    v_mov_b32_e32 v48, s62
+; GFX11-NEXT:    v_mov_b32_e32 v54, s72
+; GFX11-NEXT:    v_mov_b32_e32 v64, s60
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v70, s56 :: v_dual_mov_b32 v183, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 21
+; GFX11-NEXT:    v_mov_b32_e32 v80, s46
+; GFX11-NEXT:    v_mov_b32_e32 v18, s76
+; GFX11-NEXT:    v_mov_b32_e32 v40, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 22
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v182, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 23
+; GFX11-NEXT:    v_mov_b32_e32 v181, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 24
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v180, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 25
+; GFX11-NEXT:    v_mov_b32_e32 v178, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 26
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v179, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 27
+; GFX11-NEXT:    v_mov_b32_e32 v177, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 28
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v176, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 29
+; GFX11-NEXT:    v_mov_b32_e32 v167, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 30
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v165, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 31
+; GFX11-NEXT:    v_mov_b32_e32 v166, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v164, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 1
+; GFX11-NEXT:    v_mov_b32_e32 v163, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v162, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 3
+; GFX11-NEXT:    v_mov_b32_e32 v160, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v161, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 5
+; GFX11-NEXT:    v_mov_b32_e32 v151, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v150, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 7
+; GFX11-NEXT:    v_mov_b32_e32 v149, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_mov_b32_e32 v148, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 0
+; GFX11-NEXT:    v_readlane_b32 s1, v78, 1
+; GFX11-NEXT:    v_mov_b32_e32 v82, s0
+; GFX11-NEXT:  .LBB95_5: ; %end
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v74
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xff, v52
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 8, v82
+; GFX11-NEXT:    v_and_b32_e32 v53, 0xff, v53
+; GFX11-NEXT:    v_lshlrev_b32_e32 v82, 8, v63
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v81
+; GFX11-NEXT:    v_or_b32_e32 v52, v52, v69
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xff, v73
+; GFX11-NEXT:    v_and_b32_e32 v50, 0xff, v50
+; GFX11-NEXT:    v_and_b32_e32 v57, 0xff, v57
+; GFX11-NEXT:    v_lshlrev_b32_e32 v58, 8, v58
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v52
+; GFX11-NEXT:    v_or_b32_e32 v66, v69, v66
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v72
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v80
+; GFX11-NEXT:    v_and_b32_e32 v39, 0xff, v39
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
+; GFX11-NEXT:    v_or_b32_e32 v53, v53, v69
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xff, v62
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
+; GFX11-NEXT:    v_lshlrev_b32_e32 v67, 8, v67
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v32
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xff, v33
+; GFX11-NEXT:    v_or_b32_e32 v69, v69, v82
+; GFX11-NEXT:    v_lshlrev_b32_e32 v82, 8, v60
+; GFX11-NEXT:    v_and_b32_e32 v60, 0xff, v61
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 8, v64
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-NEXT:    v_or_b32_e32 v82, v50, v82
+; GFX11-NEXT:    v_or_b32_e32 v81, v60, v81
+; GFX11-NEXT:    v_or_b32_e32 v50, v52, v66
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v53
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 16, v69
+; GFX11-NEXT:    v_and_b32_e32 v66, 0xff, v51
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v59
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v25
+; GFX11-NEXT:    v_or_b32_e32 v51, v52, v53
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v82
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 16, v81
+; GFX11-NEXT:    v_or_b32_e32 v66, v66, v69
+; GFX11-NEXT:    v_or_b32_e32 v69, v57, v58
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v56
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xff, v47
+; GFX11-NEXT:    v_or_b32_e32 v52, v52, v53
+; GFX11-NEXT:    v_and_b32_e32 v53, 0xffff, v66
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v69
+; GFX11-NEXT:    v_or_b32_e32 v38, v38, v81
+; GFX11-NEXT:    v_or_b32_e32 v69, v82, v80
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v46
+; GFX11-NEXT:    v_and_b32_e32 v81, 0xff, v44
+; GFX11-NEXT:    v_lshlrev_b32_e32 v82, 8, v45
+; GFX11-NEXT:    v_or_b32_e32 v53, v53, v66
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v69
+; GFX11-NEXT:    v_or_b32_e32 v39, v39, v80
+; GFX11-NEXT:    v_or_b32_e32 v69, v81, v82
+; GFX11-NEXT:    v_and_b32_e32 v80, 0xff, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v43
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xff, v42
+; GFX11-NEXT:    v_or_b32_e32 v36, v38, v66
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v39
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v69
+; GFX11-NEXT:    v_or_b32_e32 v66, v80, v81
+; GFX11-NEXT:    v_or_b32_e32 v69, v82, v70
+; GFX11-NEXT:    v_and_b32_e32 v70, 0xff, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v41
+; GFX11-NEXT:    v_and_b32_e32 v81, 0xff, v183
+; GFX11-NEXT:    v_lshlrev_b32_e32 v82, 8, v40
+; GFX11-NEXT:    v_or_b32_e32 v37, v38, v39
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v66
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v69
+; GFX11-NEXT:    v_or_b32_e32 v66, v70, v80
+; GFX11-NEXT:    v_or_b32_e32 v69, v81, v82
+; GFX11-NEXT:    v_and_b32_e32 v80, 0xff, v181
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v182
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v180
+; GFX11-NEXT:    v_and_b32_e32 v66, 0xffff, v66
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
+; GFX11-NEXT:    v_or_b32_e32 v67, v80, v67
+; GFX11-NEXT:    v_or_b32_e32 v32, v32, v70
+; GFX11-NEXT:    v_or_b32_e32 v33, v33, v81
+; GFX11-NEXT:    v_or_b32_e32 v38, v38, v39
+; GFX11-NEXT:    v_or_b32_e32 v39, v66, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v67
+; GFX11-NEXT:    v_and_b32_e32 v67, 0xff, v178
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v179
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v177
+; GFX11-NEXT:    v_and_b32_e32 v80, 0xff, v176
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v167
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xff, v165
+; GFX11-NEXT:    v_lshlrev_b32_e32 v165, 8, v166
+; GFX11-NEXT:    v_or_b32_e32 v67, v67, v69
+; GFX11-NEXT:    v_or_b32_e32 v28, v28, v70
+; GFX11-NEXT:    v_or_b32_e32 v64, v80, v64
+; GFX11-NEXT:    v_or_b32_e32 v29, v29, v81
+; GFX11-NEXT:    v_or_b32_e32 v69, v82, v165
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-NEXT:    v_lshlrev_b32_e32 v67, 16, v67
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    scratch_store_b128 v0, v[50:53], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[36:39], off offset:16
+; GFX11-NEXT:    v_or_b32_e32 v36, v32, v66
+; GFX11-NEXT:    v_or_b32_e32 v37, v33, v67
+; GFX11-NEXT:    v_or_b32_e32 v38, v28, v64
+; GFX11-NEXT:    v_or_b32_e32 v39, v29, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v164
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v163
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v54
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v162
+; GFX11-NEXT:    v_and_b32_e32 v50, 0xff, v160
+; GFX11-NEXT:    v_lshlrev_b32_e32 v51, 8, v161
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 8, v151
+; GFX11-NEXT:    v_or_b32_e32 v24, v24, v28
+; GFX11-NEXT:    v_or_b32_e32 v28, v29, v32
+; GFX11-NEXT:    v_or_b32_e32 v25, v25, v33
+; GFX11-NEXT:    v_or_b32_e32 v29, v50, v51
+; GFX11-NEXT:    v_or_b32_e32 v20, v20, v52
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v150
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v48
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v149
+; GFX11-NEXT:    v_and_b32_e32 v50, 0xff, v148
+; GFX11-NEXT:    v_lshlrev_b32_e32 v51, 8, v147
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 8, v146
+; GFX11-NEXT:    v_and_b32_e32 v53, 0xff, v145
+; GFX11-NEXT:    v_lshlrev_b32_e32 v54, 8, v68
+; GFX11-NEXT:    v_or_b32_e32 v32, v32, v33
+; GFX11-NEXT:    v_or_b32_e32 v21, v21, v48
+; GFX11-NEXT:    v_or_b32_e32 v33, v50, v51
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v52
+; GFX11-NEXT:    v_or_b32_e32 v48, v53, v54
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
+; GFX11-NEXT:    v_or_b32_e32 v50, v24, v28
+; GFX11-NEXT:    v_or_b32_e32 v52, v20, v32
+; GFX11-NEXT:    v_or_b32_e32 v53, v21, v33
+; GFX11-NEXT:    v_or_b32_e32 v64, v15, v48
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v144
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v135
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v134
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 8, v133
+; GFX11-NEXT:    v_or_b32_e32 v51, v25, v29
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v132
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v65
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 8, v131
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v16, v20, v21
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v24
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v130
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 8, v129
+; GFX11-NEXT:    v_or_b32_e32 v20, v25, v28
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v29
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 8, v128
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v119
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 8, v34
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v118
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xff, v117
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 8, v116
+; GFX11-NEXT:    v_or_b32_e32 v21, v21, v24
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v25
+; GFX11-NEXT:    v_or_b32_e32 v24, v28, v29
+; GFX11-NEXT:    v_or_b32_e32 v12, v12, v32
+; GFX11-NEXT:    v_or_b32_e32 v25, v33, v34
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
+; GFX11-NEXT:    v_or_b32_e32 v65, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v66, v13, v20
+; GFX11-NEXT:    v_or_b32_e32 v67, v14, v21
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v115
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v114
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v30
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v113
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v112
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v103
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v24
+; GFX11-NEXT:    v_or_b32_e32 v12, v12, v25
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 8, v102
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v13
+; GFX11-NEXT:    v_or_b32_e32 v13, v14, v15
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v16
+; GFX11-NEXT:    v_or_b32_e32 v14, v20, v21
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v101
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v26
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v100
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 8, v97
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xff, v96
+; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 8, v22
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v24
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v99
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 8, v98
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v20
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v25
+; GFX11-NEXT:    v_or_b32_e32 v20, v26, v22
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_or_b32_e32 v16, v21, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_or_b32_e32 v13, v9, v13
+; GFX11-NEXT:    v_or_b32_e32 v14, v10, v14
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v15
+; GFX11-NEXT:    v_or_b32_e32 v9, v5, v20
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 8, v87
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v86
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v85
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v83
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v16
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v84
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v71
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v10, v15
+; GFX11-NEXT:    v_or_b32_e32 v10, v20, v18
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v16
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v49
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v55
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v35
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v31
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v27
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xff, v23
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v16
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v18
+; GFX11-NEXT:    v_or_b32_e32 v16, v20, v17
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v21
+; GFX11-NEXT:    v_or_b32_e32 v17, v22, v19
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-NEXT:    v_or_b32_e32 v10, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v1, v3, v15
+; GFX11-NEXT:    v_or_b32_e32 v2, v4, v18
+; GFX11-NEXT:    v_or_b32_e32 v3, v19, v16
+; GFX11-NEXT:    v_or_b32_e32 v4, v20, v17
+; GFX11-NEXT:    s_clause 0x5
+; GFX11-NEXT:    scratch_store_b128 v0, v[36:39], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[50:53], off offset:48
+; GFX11-NEXT:    scratch_store_b128 v0, v[64:67], off offset:64
+; GFX11-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
+; GFX11-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
+; GFX11-NEXT:    s_clause 0x12
+; GFX11-NEXT:    scratch_load_b32 v74, off, s32
+; GFX11-NEXT:    scratch_load_b32 v73, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:12
+; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:16
+; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:20
+; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:24
+; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:28
+; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:32
+; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:36
+; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:40
+; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:44
+; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:48
+; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:52
+; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:56
+; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:60
+; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:64
+; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:68
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:72
+; GFX11-NEXT:    v_readlane_b32 s104, v76, 8
+; GFX11-NEXT:    v_readlane_b32 s103, v76, 7
+; GFX11-NEXT:    v_readlane_b32 s102, v76, 6
+; GFX11-NEXT:    v_readlane_b32 s101, v76, 5
+; GFX11-NEXT:    v_readlane_b32 s100, v76, 4
+; GFX11-NEXT:    v_readlane_b32 s99, v76, 3
+; GFX11-NEXT:    v_readlane_b32 s98, v76, 2
+; GFX11-NEXT:    v_readlane_b32 s97, v76, 1
+; GFX11-NEXT:    v_readlane_b32 s96, v76, 0
+; GFX11-NEXT:    v_readlane_b32 s87, v75, 31
+; GFX11-NEXT:    v_readlane_b32 s86, v75, 30
+; GFX11-NEXT:    v_readlane_b32 s85, v75, 29
+; GFX11-NEXT:    v_readlane_b32 s84, v75, 28
+; GFX11-NEXT:    v_readlane_b32 s83, v75, 27
+; GFX11-NEXT:    v_readlane_b32 s82, v75, 26
+; GFX11-NEXT:    v_readlane_b32 s81, v75, 25
+; GFX11-NEXT:    v_readlane_b32 s80, v75, 24
+; GFX11-NEXT:    v_readlane_b32 s71, v75, 23
+; GFX11-NEXT:    v_readlane_b32 s70, v75, 22
+; GFX11-NEXT:    v_readlane_b32 s69, v75, 21
+; GFX11-NEXT:    v_readlane_b32 s68, v75, 20
+; GFX11-NEXT:    v_readlane_b32 s67, v75, 19
+; GFX11-NEXT:    v_readlane_b32 s66, v75, 18
+; GFX11-NEXT:    v_readlane_b32 s65, v75, 17
+; GFX11-NEXT:    v_readlane_b32 s64, v75, 16
+; GFX11-NEXT:    v_readlane_b32 s55, v75, 15
+; GFX11-NEXT:    v_readlane_b32 s54, v75, 14
+; GFX11-NEXT:    v_readlane_b32 s53, v75, 13
+; GFX11-NEXT:    v_readlane_b32 s52, v75, 12
+; GFX11-NEXT:    v_readlane_b32 s51, v75, 11
+; GFX11-NEXT:    v_readlane_b32 s50, v75, 10
+; GFX11-NEXT:    v_readlane_b32 s49, v75, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v75, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v75, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v75, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v75, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v75, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v75, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v75, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v75, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v75, 0
+; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:76
+; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:80
+; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:84
+; GFX11-NEXT:    scratch_load_b32 v78, off, s32 offset:88
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -212546,417 +202094,438 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64i16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:436
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:320
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v6 :: v_dual_mov_b32 v33, v0
+; GFX11-TRUE16-NEXT:    s_clause 0x1e
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:440
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:352
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:320
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, v30 :: v_dual_mov_b32 v51, v24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v28 :: v_dual_mov_b32 v55, v26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v22 :: v_dual_mov_b32 v48, v20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v18 :: v_dual_mov_b32 v39, v16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v14 :: v_dual_mov_b32 v34, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v12 :: v_dual_mov_b32 v38, v10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v6 :: v_dual_mov_b32 v33, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v2 :: v_dual_mov_b32 v32, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
 ; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:316
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v89, off, s32 offset:312
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v92, off, s32 offset:308
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v91, off, s32 offset:304
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v93, off, s32 offset:300
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:296
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:292
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v88, off, s32 offset:288
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v90, off, s32 offset:284
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:280
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:276
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:272
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:268
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:264
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:260
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:256
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:252
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:248
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:244
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:240
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:236
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:232
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:228
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:224
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:220
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:216
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:212
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:208
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:204
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:200
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:196
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v2, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v16, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v18, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v20, off, s32 offset:72
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v22, off, s32 offset:80
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v24, off, s32 offset:88
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v26, off, s32 offset:96
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v28, off, s32 offset:104
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v30, off, s32 offset:112
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:120
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:128
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:136
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:144
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:152
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:160
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:168
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:176
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:184
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:192
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:200
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:208
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:216
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:224
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:232
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:240
 ; GFX11-TRUE16-NEXT:    s_clause 0x1f
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:188
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:184
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:180
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:176
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:172
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:168
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:164
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:160
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:156
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:152
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:148
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:144
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:140
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:136
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:132
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:128
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:124
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:120
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:116
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:112
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:108
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:104
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:100
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:96
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:92
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:64
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:248
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:256
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v88, off, s32 offset:264
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v89, off, s32 offset:272
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v90, off, s32 offset:280
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v91, off, s32 offset:288
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v92, off, s32 offset:296
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v93, off, s32 offset:304
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v94, off, s32 offset:312
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:308
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:300
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:292
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:284
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:276
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:268
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:260
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:252
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:244
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:236
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:228
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:220
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:212
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:204
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:196
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:188
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:180
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:172
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:164
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:156
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:148
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:140
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:132
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:124
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:116
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:108
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:100
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:92
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:84
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:76
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:68
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:12
 ; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v86, v30 :: v_dual_mov_b32 v81, v29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v84, v28 :: v_dual_mov_b32 v83, v27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v85, v26 :: v_dual_mov_b32 v80, v24
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v69, v25 :: v_dual_mov_b32 v82, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v71, v23 :: v_dual_mov_b32 v64, v21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v67, v20 :: v_dual_mov_b32 v68, v19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v70, v18 :: v_dual_mov_b32 v55, v16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v17 :: v_dual_mov_b32 v65, v15
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v66, v14 :: v_dual_mov_b32 v49, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v12 :: v_dual_mov_b32 v54, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v11 :: v_dual_mov_b32 v36, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v8 :: v_dual_mov_b32 v37, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v7 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v32, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v101, 8, v6
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v102, 8, v8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v113, 8, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v112, 8, v12
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v130, 8, v14
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v128, 8, v16
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v134, 8, v18
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v132, 8, v20
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v161, 8, v22
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v147, 8, v24
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v166, 8, v26
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v167, 8, v28
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v180, 8, v30
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v177, 8, v31
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v42, 8, v41
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(61)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 8, v44
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(60)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v45, 8, v45
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(59)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v44, 8, v56
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(58)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v59, 8, v59
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(57)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v56, 8, v60
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(56)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v60, 8, v61
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(55)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v61, 8, v62
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(54)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v63, 8, v63
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(53)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v62, 8, v72
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(52)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v73, 8, v73
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(51)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v72, 8, v74
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(50)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v75, 8, v75
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(49)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v74, 8, v76
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(48)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v77, 8, v77
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(47)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v76, 8, v78
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(46)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v78, 8, v79
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(45)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v79, 8, v88
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(44)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v89, 8, v89
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(43)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v88, 8, v90
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(42)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v91, 8, v91
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(41)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v90, 8, v92
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(40)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v92, 8, v93
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(39)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v93, 8, v94
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB97_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v33
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v32
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v35
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v2, 0xffff, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v1, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v53
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v6, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v49
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v66
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v65
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v55
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v68
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v3, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v2, 16, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v83
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v86
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v98
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v96
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v102
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v101
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v99
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v114
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v112
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v103
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v118
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v115
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v116
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v113
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v130
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v119
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v14, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v134
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v133
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v128
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v117
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v146
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v145
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v132
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v135
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v131
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v129
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v144
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v150
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v148
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v21
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v147
+; GFX11-TRUE16-NEXT:    v_and_b32_e64 v5, 0xffff, s5
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
 ; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v149
 ; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v162
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v151
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v160
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v161
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v166
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v163
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s26, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v164
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s27, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
 ; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v165
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v178
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v177
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v68
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v64
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v4, v67
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v37
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v70
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v7, v80
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v8, v81
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v9, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v86
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v83
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v96
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v85
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v10, v97
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v11, v87
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v99
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v3, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v103
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v114
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v98
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v0, 16, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v100
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v113
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v101
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v116
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v14, v128
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v112
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v117
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v102
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v130
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v133
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v14, v132
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v0, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v148
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v118
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v129
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v161
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v2, 16, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v166
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v144
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v134
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v147
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v167
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v17, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v22
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v151
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v149
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v20, 16, v21
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v167
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v180
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v177
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v165
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v162
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v42
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v41
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v178
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v115
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v45
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v44
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v131
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v119
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v59
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v56
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v176
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v182
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v181
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v179
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v145
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v135
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v60
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v61
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v180
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v42
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v40
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v183
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v150
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v146
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v63
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v62
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v41
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v46
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v43
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v44
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v163
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v160
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v73
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v72
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v45
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v47
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v56
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v176
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v164
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v75
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v74
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v57
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v62
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v61
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v59
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v179
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v77
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v76
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v60
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v74
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v73
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v63
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v183
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v182
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v78
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v79
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v72
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v77
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v76
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v75
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v43
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v40
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v89
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v88
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v78
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v90
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v88
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v79
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v47
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v46
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v91
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v90
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v89
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v93
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v91
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v92
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v58
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v57
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v92
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v93
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v1, 16, v0
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
@@ -213003,396 +202572,344 @@ define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i3
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v92
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v93
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(37)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v58
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v57
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(35)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v47
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v90
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v79
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v92, v0
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v91
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v88
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v78
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v77
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v7, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v46
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v93, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v91, v2
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(33)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v43
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v5, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v76
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v75
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v74
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v72
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v73
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v40
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v90, v3
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(31)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v183
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v182
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v63
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(59)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v62
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v61
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v59
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v59, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v60, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v57
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(55)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v58
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v57, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v56
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v56, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v47
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v89, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(29)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v181
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v88, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v181, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v78, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v79, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v179
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v182, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v179, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v77, v3
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(27)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v176
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v45
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(51)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v46
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v45, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v44
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v43
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v41
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(47)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v42
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v164
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(25)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v163
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v163, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v76, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v160
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v75, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v160, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v74, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v73, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v1
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(23)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v150
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v40
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v183
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(43)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v182
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v180
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v181
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v179
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(39)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v178
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v176
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v72, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v146
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v177
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v167
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v167, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v176, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v165
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(35)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v166
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v165, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v164
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v164, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v163
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(21)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v145
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v135
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v63, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(19)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v131
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v62, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v131, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v60, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v61, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v119
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v135, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v119, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v59, v3
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(17)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v178
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v161
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v162
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v161, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v160
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v151
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v149
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v150
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v115
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v165
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v115, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v56, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v162
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v45, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v145, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v44, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v42, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v1
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v151
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v148
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v147
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(23)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v146
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v144
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v145
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v135
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(19)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v134
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v132
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v41, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v149
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v133
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v131
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v131, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v132, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v129
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v130
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v129, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v128
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v128, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v119
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v117
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(11)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v118
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v117, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v116
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v115
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v113
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v148
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v144
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v180, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(9)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v133
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v177, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v166, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v167, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v133, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v129
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v129, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v144, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v161, v3
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v114
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v4, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v112
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v5, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v103
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v102
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v101
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v99
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v86
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v118
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v117
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(5)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v116
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v116, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v147, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v114
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v96
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v55, 3, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v99
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v114, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v132, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v130, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v103
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v98
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v54
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v39
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 3, v33
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v113, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v128, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v100
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v101, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v102, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v98
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v96
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v134, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v97, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v87
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v85
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v84
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v83
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v81
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v82
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v87, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v51
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v86, v4
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v69
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v70
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v70, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v69, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v68
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v67
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v66
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v64
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v85, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v84, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v49
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v83, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v48
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v65
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v64, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v52
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v82, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v38
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v81, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v71, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v80, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v37
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v70, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v36
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, 3, v51
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v53, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v52, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v49
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v50
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 3, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v38, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v36, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v37, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v33
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v35
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v69, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v112, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v68, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v67, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v66, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v33
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v34, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v32
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v65, v8
 ; GFX11-TRUE16-NEXT:    v_and_b32_e64 v8, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v64
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v64, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v7, 16, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v32, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v51
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v38
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v53, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v49, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v69
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v33, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v34, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v50, 16, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v132
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v3, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v116
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v129
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v18
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v114, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v144, 16, v33
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v128
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v129, 16, v33
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v176
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v117, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v164
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v165, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v161, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v56
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v14, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v115
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v135
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v131
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v23
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v27
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v145, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v119, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v163
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v182
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v181
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v28
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v2, 16, v0
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v131, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v167, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v45, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v57, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v59, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v133, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v160, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v179, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v36
 ; GFX11-TRUE16-NEXT:  .LBB97_3: ; %end
-; GFX11-TRUE16-NEXT:    s_clause 0x1d
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:320
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:324
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:328
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:332
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:336
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:340
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:344
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:348
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:352
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:356
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:360
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:364
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:368
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:372
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:376
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:380
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:384
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:388
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:392
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:396
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:400
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:404
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:408
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:412
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:416
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:420
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:424
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:428
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:432
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    s_clause 0x1e
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:320
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:324
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:328
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:332
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:336
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:340
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:344
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:348
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:352
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:356
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:360
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:364
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:368
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:372
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:376
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:380
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:384
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:388
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:392
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:396
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:400
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:404
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:408
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:412
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:416
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:420
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:424
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:428
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:432
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:436
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:440
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ; GFX11-TRUE16-NEXT:  .LBB97_4:
@@ -223614,2095 +213131,1034 @@ define inreg <128 x i8> @bitcast_v64i16_to_v128i8_scalar(<64 x i16> inreg %a, i3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v64i16_to_v128i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:88
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s30, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s96, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s97, 1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s98, 2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s99, 3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s100, 4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s101, 5
-; GFX11-TRUE16-NEXT:    s_mov_b32 vcc_hi, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_clause 0x12
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:72
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s102, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s103, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s104, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s50, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s51, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s52, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s53, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s54, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s55, 15
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s64, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s65, 17
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s66, 18
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s67, 19
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s68, 20
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s69, 21
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s70, 22
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s71, 23
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s80, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s81, 25
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s82, 26
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s83, 27
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s84, 28
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s85, 29
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s86, 30
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s87, 31
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB99_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[0:1], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s5, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s5, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s5, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s4, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s7, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s7, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s51, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s52, s6, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s53, s9, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s54, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s55, s9, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s64, s8, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s65, s8, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s66, s11, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s67, s11, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s68, s11, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s69, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s70, s10, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s71, s13, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s80, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s81, s13, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s82, s12, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 31
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s83, s12, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s84, s15, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s85, s15, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 30
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s86, s15, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s87, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s96, s14, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 29
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s97, s41, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s98, s41, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s41, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 28
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s100, s40, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s101, s40, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s29, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 27
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s29, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s28, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 26
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[62:63], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[72:73], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 25
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 23
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[88:89], s[10:11], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[14:15], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 22
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[40:41], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 21
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 20
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 19
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s18, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 18
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s18, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 17
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 15
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 14
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 13
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s16, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 24
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 11
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 10
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 9
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s2, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s2, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 7
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 5
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s0, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s0, 8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s74, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s75, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[4:5], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_hi
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB99_4
-; GFX11-TRUE16-NEXT:  .LBB99_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s6, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v51, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v50, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v36, s19, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s21, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v28, s27, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v27, s26, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v35, s18, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v22, s29, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, s28, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, s41, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s40, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s15, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s4, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v55, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v54, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v49, s17, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v48, s16, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s23, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v32, s25, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v31, s24, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s22, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[23:24], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[27:28]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[35:36]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[80:81], 24, v[50:51]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[24:25], 24, v[9:10]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[31:32]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[48:49]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[81:82], 24, v[54:55]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[5:6]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[25:26], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[29:30], 24, v[13:14]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[17:18]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[21:22]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 24, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v164, 8, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v165, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v166, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v167, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v176, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v177, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v178, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v179, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v180, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v183, 24, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v40, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v41, 8, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v44, 24, v49
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v45, 16, v49
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v46, 8, v49
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v48
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v57, 24, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v58, 16, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v51
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v60, 16, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v61, 8, v50
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v62, 24, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v63, 16, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v72, 8, v55
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v54
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v54
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 24, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 24, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 24, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 24, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 8, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v21
-; GFX11-TRUE16-NEXT:    s_branch .LBB99_5
-; GFX11-TRUE16-NEXT:  .LBB99_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr104_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr103_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr102_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr101_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr99_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr98_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr97_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr96_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr87_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr86_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr85_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr84_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s43, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 8
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 9
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 10
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 11
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 12
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 13
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 14
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 15
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 17
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 18
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 19
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 20
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 21
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 22
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 23
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 24
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 25
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 26
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 27
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 28
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 29
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 30
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v78, s42, 31
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 0
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 1
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 2
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 3
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 4
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 5
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 6
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 7
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 8
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB99_2
-; GFX11-TRUE16-NEXT:  .LBB99_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v74.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v73.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v72.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v63.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v61.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v150.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v59.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v149.l, s34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v144.l, s104
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v135.l, s103
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v134.l, s102
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v133.l, s101
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v132.l, s100
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v131.l, s99
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v130.l, s98
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v129.l, s97
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.l, s96
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v119.l, s87
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v118.l, s86
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v56.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v117.l, s85
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v116.l, s84
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v115.l, s83
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v114.l, s82
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v113.l, s81
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v112.l, s80
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v103.l, s71
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v102.l, s70
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v46.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v101.l, s69
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v100.l, s68
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v99.l, s67
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v98.l, s66
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v97.l, s65
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v96.l, s64
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, s55
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v86.l, s54
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v44.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.l, s53
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.l, s52
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v83.l, s51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v82.l, s50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v43.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v71.l, s49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v145.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v41.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s72
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v70.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v80.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s74
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v183.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s76
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s78
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s88
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s90
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v182.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s92
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, s94
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v68.l, s30
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v181.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v180.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 25
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v179.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v178.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 27
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v177.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v176.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 29
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v167.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 30
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v166.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 31
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v165.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v164.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v163.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v162.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v161.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v160.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v151.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v148.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v147.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v146.l, s0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v78, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s1, v78, 3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v81.l, s0
-; GFX11-TRUE16-NEXT:  .LBB99_5: ; %end
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v74
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xff, v54
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xff, v55
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v62
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, v54, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v73
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xff, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v60, 0xff, v60
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v69, v66
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v72
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v80, v60, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v57, 8, v57
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, v55, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v63
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xff, v49
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v54, v54, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, v69, v81
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v61
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v50, v81
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v59
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v81, 0xff, v58
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v64
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, v55, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v51, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, v81, v57
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v56
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v81, 0xff, v47
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v56, v50, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, v48, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v81, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v46
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xff, v45
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v44
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v57, v50, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, v49, v69
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v70, v80
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v43
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v42
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, v48, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, v35, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v69, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v41
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v40
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v183
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, v49, v50
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v36, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v67, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v181
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v182
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v64, v67, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v180
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v35, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v36, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xff, v179
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v178
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v177
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v176
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v167
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v36, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v52, v67, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[54:57], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[48:51], off offset:16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v164
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xff, v163
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v162
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xff, v161
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v160
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v151
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v66, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v31, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, v36, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v32, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v49, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v27, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xff, v148
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v147
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xff, v146
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v145
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v150
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xff, v149
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, v38, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v28, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v49, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v51
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, v52, v54
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, v31, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, v27, v37
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, v28, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, v21, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v144
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v135
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v134
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v133
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v32, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v132
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 8, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v131
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v27, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v17, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v130
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v129
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v32, v49
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v128
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xff, v119
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v28, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v118
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xff, v117
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v116
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v49, v33
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v50
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v51, v52
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, v21, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v50, v17, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, v18, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v115
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v114
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v113
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v102
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v112
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v103
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v18, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v101
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v96
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v27, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v100
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xff, v99
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v98
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v31, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v27, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v14, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v9, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v86
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v85
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xff, v166
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v165
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v84
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v14, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v24, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v64, v70, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v53
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v24, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v26, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v67, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v23, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v24, v20
-; GFX11-TRUE16-NEXT:    s_clause 0x5
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[35:38], off offset:48
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[48:51], off offset:64
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[15:18], off offset:80
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[11:14], off offset:96
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:112
-; GFX11-TRUE16-NEXT:    s_clause 0x12
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:64
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:68
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:72
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s104, v76, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s103, v76, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s102, v76, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s101, v76, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s100, v76, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s99, v76, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s98, v76, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s97, v76, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s96, v76, 0
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s87, v75, 31
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s86, v75, 30
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s85, v75, 29
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s84, v75, 28
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s83, v75, 27
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s82, v75, 26
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s81, v75, 25
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s80, v75, 24
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s71, v75, 23
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s70, v75, 22
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s69, v75, 21
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s68, v75, 20
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v75, 19
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v75, 18
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v75, 17
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v75, 16
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v75, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v75, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v75, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v75, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v75, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v75, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v75, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v75, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v75, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v75, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v75, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v75, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v75, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v75, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v75, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v75, 0
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:76
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:80
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:84
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:88
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v64i16_to_v128i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:88
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s96, 0
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s97, 1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s98, 2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s99, 3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s100, 4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s101, 5
-; GFX11-FAKE16-NEXT:    s_mov_b32 s99, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_clause 0x12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:72
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s102, 6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s103, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s104, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s49, 9
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s50, 10
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s51, 11
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s52, 12
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s53, 13
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s54, 14
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s55, 15
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s64, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s65, 17
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s66, 18
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s67, 19
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s68, 20
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s69, 21
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s70, 22
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s71, 23
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s80, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s81, 25
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s82, 26
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s83, 27
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s84, 28
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s85, 29
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s86, 30
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s87, 31
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB99_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[74:75], s[0:1], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s5, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s5, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s5, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s4, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s7, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s50, s7, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s51, s6, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s52, s9, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s53, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s55, s9, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s54, s8, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s64, s8, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s65, s11, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s66, s11, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s68, s11, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s67, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s69, s10, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s70, s13, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s71, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s81, s13, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s80, s12, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 31
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s82, s12, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s83, s15, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s84, s15, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 30
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s86, s15, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s85, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s87, s14, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 29
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s96, s41, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s97, s41, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s41, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 28
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s98, s40, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s101, s40, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s29, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 27
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s29, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s28, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 26
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[72:73], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 25
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 23
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[88:89], s[10:11], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[92:93], s[14:15], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 22
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[40:41], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 21
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 20
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 19
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 18
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 17
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 15
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 14
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 13
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 24
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 11
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 10
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 9
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s2, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s2, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 7
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 6
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 5
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 4
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s0, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s0, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 2
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s74, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s75, 1
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[74:75], s[4:5], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s99
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB99_4
-; GFX11-FAKE16-NEXT:  .LBB99_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v39, s17, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v38, s16, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v2, s5, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v1, s4, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v51, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v50, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v33, s21, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v29, s23, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v28, s22, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v32, s20, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v16, s29, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v15, s28, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v14, s41, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v13, s40, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v12, s15, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v10, s13, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v8, s11, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v6, s9, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v5, s8, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v4, s7, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v3, s6, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v53, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v52, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v37, s19, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v25, s25, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v21, s27, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v20, s26, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v24, s24, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v36, s18, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[80:81], 24, v[38:39]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[28:29]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[32:33]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[81:82], 24, v[50:51]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[20:21]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[24:25]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[36:37]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[82:83], 24, v[52:53]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[18:19], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[7:8]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[30:31], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[11:12]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[15:16]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 8, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 8, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 8, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 8, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 16, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 8, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 8, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v51
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v51
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v51
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v50
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v50
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 16, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 8, v53
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v52
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v52
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 24, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 24, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 24, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v15
-; GFX11-FAKE16-NEXT:    s_branch .LBB99_5
-; GFX11-FAKE16-NEXT:  .LBB99_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    s_mov_b32 s99, -1
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr104
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr103
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr102
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr101
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr98
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr100
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr97
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr96
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr87
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr85
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr86
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr84
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s42, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v78, s43, 1
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    s_branch .LBB99_2
-; GFX11-FAKE16-NEXT:  .LBB99_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s0 :: v_dual_mov_b32 v53, s1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v71, s50
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v74, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s12 :: v_dual_mov_b32 v10, s13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v73, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v55, s48
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v72, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v49, s39
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s6 :: v_dual_mov_b32 v4, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s4 :: v_dual_mov_b32 v2, s5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v62, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, s2 :: v_dual_mov_b32 v51, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s16 :: v_dual_mov_b32 v39, s17
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v63, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v35, s38 :: v_dual_mov_b32 v36, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v37, s19 :: v_dual_mov_b32 v32, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v33, s21 :: v_dual_mov_b32 v60, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s22 :: v_dual_mov_b32 v29, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v61, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s26 :: v_dual_mov_b32 v21, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v146, s42 :: v_dual_mov_b32 v145, s104
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v59, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v144, vcc_hi :: v_dual_mov_b32 v135, s103
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v134, s102 :: v_dual_mov_b32 v133, s101
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v57, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v31, s36 :: v_dual_mov_b32 v132, s98
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v131, s100 :: v_dual_mov_b32 v130, s97
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v129, s96 :: v_dual_mov_b32 v58, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s37 :: v_dual_mov_b32 v128, s87
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v119, s85 :: v_dual_mov_b32 v118, s86
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v117, s84 :: v_dual_mov_b32 v56, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v116, s83 :: v_dual_mov_b32 v115, s82
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v114, s80 :: v_dual_mov_b32 v113, s81
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v47, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s35 :: v_dual_mov_b32 v112, s71
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v103, s70 :: v_dual_mov_b32 v102, s69
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v101, s67 :: v_dual_mov_b32 v46, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s34 :: v_dual_mov_b32 v100, s68
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v99, s66 :: v_dual_mov_b32 v98, s65
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v97, s64 :: v_dual_mov_b32 v44, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v96, s54 :: v_dual_mov_b32 v87, s55
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s53 :: v_dual_mov_b32 v85, s52
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v45, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v84, s51 :: v_dual_mov_b32 v83, s49
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v147, s43 :: v_dual_mov_b32 v22, s78
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v43, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v67, s58 :: v_dual_mov_b32 v26, s88
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v81, s44 :: v_dual_mov_b32 v30, s90
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v42, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s74 :: v_dual_mov_b32 v34, s92
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v65, s94 :: v_dual_mov_b32 v68, s30
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v41, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 20
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v48, s62
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v54, s72
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v64, s60
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, s56 :: v_dual_mov_b32 v183, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 21
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v80, s46
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v18, s76
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v40, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 22
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v182, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 23
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v181, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 24
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v180, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 25
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v178, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 26
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v179, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v177, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 28
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v176, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 29
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v167, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 30
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v165, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 31
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v166, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v164, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v163, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v162, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v160, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v161, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v151, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v150, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 7
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v149, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v148, s0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v78, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s1, v78, 1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v82, s0
-; GFX11-FAKE16-NEXT:  .LBB99_5: ; %end
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v74
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xff, v52
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xff, v53
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v63
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v52, v69
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xff, v73
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xff, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v57, 0xff, v57
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v58, 8, v58
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v69, v66
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v72
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v80
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v53, v69
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xff, v62
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v67
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v69, v82
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v60
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v60, 0xff, v61
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v64
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, v50, v82
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v81, v60, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, v52, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v53
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v69
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xff, v51
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v59
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, v52, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v82
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v53, 16, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v66, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v57, v58
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v56
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xff, v47
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v52, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xffff, v66
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, v38, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v82, v80
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v46
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v81, 0xff, v44
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v45
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v53, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, v39, v80
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v81, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v80, 0xff, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v43
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xff, v42
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v38, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v39
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v80, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v82, v70
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v70, 0xff, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v41
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v81, 0xff, v183
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v40
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v38, v39
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v66
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v70, v80
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v81, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v80, 0xff, v181
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v182
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v180
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xffff, v66
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, v80, v67
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v32, v70
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v33, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, v38, v39
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, v66, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v67
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v67, 0xff, v178
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v179
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v177
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v80, 0xff, v176
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v167
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xff, v165
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v165, 8, v166
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, v67, v69
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v70
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, v80, v64
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v29, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, v82, v165
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v67
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[50:53], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[36:39], off offset:16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v32, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, v33, v67
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, v28, v64
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, v29, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v164
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v163
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v54
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v162
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xff, v160
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v161
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v151
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v25, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v50, v51
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v52
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v150
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v149
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xff, v148
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v147
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v146
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xff, v145
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v68
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v32, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v50, v51
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, v53, v54
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, v24, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, v20, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, v21, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, v15, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v144
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v135
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v134
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v133
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, v25, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v132
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v65
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v131
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v130
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v129
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v25, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v128
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v119
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v118
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v117
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v116
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v28, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v33, v34
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v13, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, v14, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v115
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v114
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v30
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v113
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v112
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v103
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v24
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v102
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v20, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v101
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v100
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v97
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v96
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v99
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v98
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v26, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v21, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v9, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v10, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v5, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v87
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v86
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v85
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v83
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v84
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v71
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v10, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v20, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v49
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v55
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v31
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v23
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v22, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v19, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v20, v17
-; GFX11-FAKE16-NEXT:    s_clause 0x5
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[36:39], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[50:53], off offset:48
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[64:67], off offset:64
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
-; GFX11-FAKE16-NEXT:    s_clause 0x12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:4
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:8
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:12
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:16
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:20
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:24
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:28
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:32
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:36
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:40
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:44
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:48
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:52
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:56
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:60
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:64
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:68
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:72
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s104, v76, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s103, v76, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s102, v76, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s101, v76, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s100, v76, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s99, v76, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s98, v76, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s97, v76, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s96, v76, 0
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s87, v75, 31
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s86, v75, 30
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s85, v75, 29
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s84, v75, 28
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s83, v75, 27
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s82, v75, 26
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s81, v75, 25
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s80, v75, 24
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s71, v75, 23
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s70, v75, 22
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s69, v75, 21
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s68, v75, 20
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v75, 19
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v75, 18
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v75, 17
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v75, 16
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v75, 15
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v75, 14
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v75, 13
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v75, 12
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v75, 11
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v75, 10
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v75, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v75, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v75, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v75, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v75, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v75, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v75, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v75, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v75, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v75, 0
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:76
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:80
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:84
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:88
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v64i16_to_v128i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:76
+; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:80
+; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:84
+; GFX11-NEXT:    scratch_store_b32 off, v78, s32 offset:88
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v75, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v76, s96, 0
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v15
+; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s41, v2
+; GFX11-NEXT:    v_writelane_b32 v75, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v76, s97, 1
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s15, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
+; GFX11-NEXT:    v_writelane_b32 v75, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v76, s98, 2
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v8
+; GFX11-NEXT:    v_writelane_b32 v75, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v76, s99, 3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
+; GFX11-NEXT:    v_writelane_b32 v75, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v76, s100, 4
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v14
+; GFX11-NEXT:    v_writelane_b32 v75, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v76, s101, 5
+; GFX11-NEXT:    s_mov_b32 s99, 0
+; GFX11-NEXT:    s_and_b32 s42, vcc_lo, exec_lo
+; GFX11-NEXT:    s_clause 0x12
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:72
+; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:68
+; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:64
+; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:60
+; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:56
+; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:52
+; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:48
+; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:44
+; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:40
+; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:36
+; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:32
+; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:28
+; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:24
+; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:20
+; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:16
+; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:12
+; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:8
+; GFX11-NEXT:    scratch_store_b32 off, v73, s32 offset:4
+; GFX11-NEXT:    scratch_store_b32 off, v74, s32
+; GFX11-NEXT:    v_writelane_b32 v75, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v76, s102, 6
+; GFX11-NEXT:    ; implicit-def: $vgpr78 : SGPR spill to VGPR lane
+; GFX11-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
+; GFX11-NEXT:    v_writelane_b32 v75, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v76, s103, 7
+; GFX11-NEXT:    v_writelane_b32 v75, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v76, s104, 8
+; GFX11-NEXT:    v_writelane_b32 v75, s49, 9
+; GFX11-NEXT:    v_writelane_b32 v75, s50, 10
+; GFX11-NEXT:    v_writelane_b32 v75, s51, 11
+; GFX11-NEXT:    v_writelane_b32 v75, s52, 12
+; GFX11-NEXT:    v_writelane_b32 v75, s53, 13
+; GFX11-NEXT:    v_writelane_b32 v75, s54, 14
+; GFX11-NEXT:    v_writelane_b32 v75, s55, 15
+; GFX11-NEXT:    v_writelane_b32 v75, s64, 16
+; GFX11-NEXT:    v_writelane_b32 v75, s65, 17
+; GFX11-NEXT:    v_writelane_b32 v75, s66, 18
+; GFX11-NEXT:    v_writelane_b32 v75, s67, 19
+; GFX11-NEXT:    v_writelane_b32 v75, s68, 20
+; GFX11-NEXT:    v_writelane_b32 v75, s69, 21
+; GFX11-NEXT:    v_writelane_b32 v75, s70, 22
+; GFX11-NEXT:    v_writelane_b32 v75, s71, 23
+; GFX11-NEXT:    v_writelane_b32 v75, s80, 24
+; GFX11-NEXT:    v_writelane_b32 v75, s81, 25
+; GFX11-NEXT:    v_writelane_b32 v75, s82, 26
+; GFX11-NEXT:    v_writelane_b32 v75, s83, 27
+; GFX11-NEXT:    v_writelane_b32 v75, s84, 28
+; GFX11-NEXT:    v_writelane_b32 v75, s85, 29
+; GFX11-NEXT:    v_writelane_b32 v75, s86, 30
+; GFX11-NEXT:    v_writelane_b32 v75, s87, 31
+; GFX11-NEXT:    s_cbranch_scc0 .LBB99_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b64 s[74:75], s[0:1], 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s34, s5, 24
+; GFX11-NEXT:    s_lshr_b32 s35, s5, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 7
+; GFX11-NEXT:    s_lshr_b32 s42, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s37, s5, 8
+; GFX11-NEXT:    s_lshr_b32 s36, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s38, s4, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 6
+; GFX11-NEXT:    s_lshr_b32 s42, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s7, 24
+; GFX11-NEXT:    s_lshr_b32 s48, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s50, s7, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 5
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s49, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s51, s6, 8
+; GFX11-NEXT:    s_lshr_b32 s52, s9, 24
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 4
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s53, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s55, s9, 8
+; GFX11-NEXT:    s_lshr_b32 s54, s8, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s64, s8, 8
+; GFX11-NEXT:    s_lshr_b32 s65, s11, 24
+; GFX11-NEXT:    s_lshr_b32 s66, s11, 16
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 2
+; GFX11-NEXT:    s_lshr_b32 s42, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s68, s11, 8
+; GFX11-NEXT:    s_lshr_b32 s67, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s69, s10, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 1
+; GFX11-NEXT:    s_lshr_b32 s42, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s70, s13, 24
+; GFX11-NEXT:    s_lshr_b32 s71, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s81, s13, 8
+; GFX11-NEXT:    v_writelane_b32 v77, s42, 0
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s80, s12, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 31
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s82, s12, 8
+; GFX11-NEXT:    s_lshr_b32 s83, s15, 24
+; GFX11-NEXT:    s_lshr_b32 s84, s15, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 30
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s86, s15, 8
+; GFX11-NEXT:    s_lshr_b32 s85, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s87, s14, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 29
+; GFX11-NEXT:    s_lshr_b32 s42, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s96, s41, 24
+; GFX11-NEXT:    s_lshr_b32 s97, s41, 16
+; GFX11-NEXT:    s_lshr_b32 s100, s41, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 28
+; GFX11-NEXT:    s_lshr_b32 s42, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s98, s40, 16
+; GFX11-NEXT:    s_lshr_b32 s101, s40, 8
+; GFX11-NEXT:    s_lshr_b32 s102, s29, 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 27
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s103, s29, 16
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s29, 8
+; GFX11-NEXT:    s_lshr_b32 s104, s28, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 26
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 16
+; GFX11-NEXT:    s_lshr_b64 s[62:63], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[72:73], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 25
+; GFX11-NEXT:    s_lshr_b32 s42, s21, 8
+; GFX11-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s20, 16
+; GFX11-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
+; GFX11-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 23
+; GFX11-NEXT:    s_lshr_b32 s42, s20, 8
+; GFX11-NEXT:    s_lshr_b64 s[88:89], s[10:11], 24
+; GFX11-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
+; GFX11-NEXT:    s_lshr_b64 s[92:93], s[14:15], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 22
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 24
+; GFX11-NEXT:    s_lshr_b64 s[94:95], s[40:41], 24
+; GFX11-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 21
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 20
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 19
+; GFX11-NEXT:    s_lshr_b32 s42, s18, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 18
+; GFX11-NEXT:    s_lshr_b32 s42, s18, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 17
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 15
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 14
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 13
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 12
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 24
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 11
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 10
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 9
+; GFX11-NEXT:    s_lshr_b32 s42, s2, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s2, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 7
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 6
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 5
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 4
+; GFX11-NEXT:    s_lshr_b32 s42, s0, 16
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 3
+; GFX11-NEXT:    s_lshr_b32 s42, s0, 8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 2
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 8
+; GFX11-NEXT:    v_writelane_b32 v78, s74, 0
+; GFX11-NEXT:    v_writelane_b32 v78, s75, 1
+; GFX11-NEXT:    s_lshr_b64 s[74:75], s[4:5], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s99
+; GFX11-NEXT:    s_cbranch_vccnz .LBB99_4
+; GFX11-NEXT:  .LBB99_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v39, s17, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v38, s16, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v2, s5, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v1, s4, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v51, s3, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v50, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v33, s21, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v29, s23, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v28, s22, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v32, s20, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v16, s29, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v15, s28, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v14, s41, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v13, s40, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v12, s15, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v10, s13, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v8, s11, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v6, s9, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v5, s8, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v4, s7, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v3, s6, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v53, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v52, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v37, s19, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v25, s25, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v21, s27, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v20, s26, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v24, s24, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v36, s18, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_lshrrev_b64 v[80:81], 24, v[38:39]
+; GFX11-NEXT:    v_lshrrev_b64 v[64:65], 24, v[28:29]
+; GFX11-NEXT:    v_lshrrev_b64 v[67:68], 24, v[32:33]
+; GFX11-NEXT:    v_lshrrev_b64 v[81:82], 24, v[50:51]
+; GFX11-NEXT:    v_lshrrev_b64 v[17:18], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[48:49], 24, v[20:21]
+; GFX11-NEXT:    v_lshrrev_b64 v[54:55], 24, v[24:25]
+; GFX11-NEXT:    v_lshrrev_b64 v[70:71], 24, v[36:37]
+; GFX11-NEXT:    v_lshrrev_b64 v[82:83], 24, v[52:53]
+; GFX11-NEXT:    v_lshrrev_b64 v[18:19], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[22:23], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b64 v[26:27], 24, v[7:8]
+; GFX11-NEXT:    v_lshrrev_b64 v[30:31], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[34:35], 24, v[11:12]
+; GFX11-NEXT:    v_lshrrev_b64 v[65:66], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b64 v[68:69], 24, v[15:16]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v147, 24, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v148, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v149, 8, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v150, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v151, 8, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v161, 24, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v160, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v162, 8, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v163, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v164, 8, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v166, 24, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v165, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v167, 8, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v176, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v177, 8, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v179, 24, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v178, 16, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v180, 8, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v181, 16, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v182, 8, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v40, 24, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v183, 16, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v41, 8, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v42, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v43, 8, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v45, 24, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v44, 16, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v46, 8, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v47, 16, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v56, 8, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v58, 24, v51
+; GFX11-NEXT:    v_lshrrev_b32_e32 v57, 16, v51
+; GFX11-NEXT:    v_lshrrev_b32_e32 v59, 8, v51
+; GFX11-NEXT:    v_lshrrev_b32_e32 v61, 16, v50
+; GFX11-NEXT:    v_lshrrev_b32_e32 v60, 8, v50
+; GFX11-NEXT:    v_lshrrev_b32_e32 v63, 24, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v62, 16, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v72, 8, v53
+; GFX11-NEXT:    v_lshrrev_b32_e32 v73, 16, v52
+; GFX11-NEXT:    v_lshrrev_b32_e32 v74, 8, v52
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 24, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 24, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 8, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v97, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v98, 24, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v99, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v100, 8, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v101, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v102, 8, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v103, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v112, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v113, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v114, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v115, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v116, 24, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v117, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v118, 8, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v119, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v128, 8, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v129, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v130, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v131, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v132, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v133, 8, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v134, 24, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v135, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v144, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v145, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v146, 8, v15
+; GFX11-NEXT:    s_branch .LBB99_5
+; GFX11-NEXT:  .LBB99_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    s_mov_b32 s99, -1
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr104
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $sgpr103
+; GFX11-NEXT:    ; implicit-def: $sgpr102
+; GFX11-NEXT:    ; implicit-def: $sgpr101
+; GFX11-NEXT:    ; implicit-def: $sgpr98
+; GFX11-NEXT:    ; implicit-def: $sgpr100
+; GFX11-NEXT:    ; implicit-def: $sgpr97
+; GFX11-NEXT:    ; implicit-def: $sgpr96
+; GFX11-NEXT:    ; implicit-def: $sgpr87
+; GFX11-NEXT:    ; implicit-def: $sgpr85
+; GFX11-NEXT:    ; implicit-def: $sgpr86
+; GFX11-NEXT:    ; implicit-def: $sgpr84
+; GFX11-NEXT:    ; implicit-def: $sgpr83
+; GFX11-NEXT:    ; implicit-def: $sgpr82
+; GFX11-NEXT:    ; implicit-def: $sgpr80
+; GFX11-NEXT:    ; implicit-def: $sgpr81
+; GFX11-NEXT:    ; implicit-def: $sgpr71
+; GFX11-NEXT:    ; implicit-def: $sgpr70
+; GFX11-NEXT:    ; implicit-def: $sgpr69
+; GFX11-NEXT:    ; implicit-def: $sgpr67
+; GFX11-NEXT:    ; implicit-def: $sgpr68
+; GFX11-NEXT:    ; implicit-def: $sgpr66
+; GFX11-NEXT:    ; implicit-def: $sgpr65
+; GFX11-NEXT:    ; implicit-def: $sgpr64
+; GFX11-NEXT:    ; implicit-def: $sgpr54
+; GFX11-NEXT:    ; implicit-def: $sgpr55
+; GFX11-NEXT:    ; implicit-def: $sgpr53
+; GFX11-NEXT:    ; implicit-def: $sgpr52
+; GFX11-NEXT:    ; implicit-def: $sgpr51
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr50
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    v_writelane_b32 v78, s42, 0
+; GFX11-NEXT:    v_writelane_b32 v78, s43, 1
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; kill: killed $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    s_branch .LBB99_2
+; GFX11-NEXT:  .LBB99_4:
+; GFX11-NEXT:    v_dual_mov_b32 v52, s0 :: v_dual_mov_b32 v53, s1
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 2
+; GFX11-NEXT:    v_mov_b32_e32 v71, s50
+; GFX11-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v74, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 3
+; GFX11-NEXT:    v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
+; GFX11-NEXT:    v_dual_mov_b32 v9, s12 :: v_dual_mov_b32 v10, s13
+; GFX11-NEXT:    v_mov_b32_e32 v73, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 4
+; GFX11-NEXT:    v_mov_b32_e32 v55, s48
+; GFX11-NEXT:    v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
+; GFX11-NEXT:    v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_mov_b32_e32 v72, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 5
+; GFX11-NEXT:    v_mov_b32_e32 v49, s39
+; GFX11-NEXT:    v_dual_mov_b32 v3, s6 :: v_dual_mov_b32 v4, s7
+; GFX11-NEXT:    v_dual_mov_b32 v1, s4 :: v_dual_mov_b32 v2, s5
+; GFX11-NEXT:    v_mov_b32_e32 v62, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 6
+; GFX11-NEXT:    v_dual_mov_b32 v50, s2 :: v_dual_mov_b32 v51, s3
+; GFX11-NEXT:    v_dual_mov_b32 v38, s16 :: v_dual_mov_b32 v39, s17
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v63, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 7
+; GFX11-NEXT:    v_dual_mov_b32 v35, s38 :: v_dual_mov_b32 v36, s18
+; GFX11-NEXT:    v_dual_mov_b32 v37, s19 :: v_dual_mov_b32 v32, s20
+; GFX11-NEXT:    v_dual_mov_b32 v33, s21 :: v_dual_mov_b32 v60, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 8
+; GFX11-NEXT:    v_dual_mov_b32 v28, s22 :: v_dual_mov_b32 v29, s23
+; GFX11-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v61, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 9
+; GFX11-NEXT:    v_dual_mov_b32 v20, s26 :: v_dual_mov_b32 v21, s27
+; GFX11-NEXT:    v_dual_mov_b32 v146, s42 :: v_dual_mov_b32 v145, s104
+; GFX11-NEXT:    v_mov_b32_e32 v59, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 10
+; GFX11-NEXT:    v_dual_mov_b32 v144, vcc_hi :: v_dual_mov_b32 v135, s103
+; GFX11-NEXT:    v_dual_mov_b32 v134, s102 :: v_dual_mov_b32 v133, s101
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v57, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 11
+; GFX11-NEXT:    v_dual_mov_b32 v31, s36 :: v_dual_mov_b32 v132, s98
+; GFX11-NEXT:    v_dual_mov_b32 v131, s100 :: v_dual_mov_b32 v130, s97
+; GFX11-NEXT:    v_dual_mov_b32 v129, s96 :: v_dual_mov_b32 v58, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 12
+; GFX11-NEXT:    v_dual_mov_b32 v27, s37 :: v_dual_mov_b32 v128, s87
+; GFX11-NEXT:    v_dual_mov_b32 v119, s85 :: v_dual_mov_b32 v118, s86
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v117, s84 :: v_dual_mov_b32 v56, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 13
+; GFX11-NEXT:    v_dual_mov_b32 v116, s83 :: v_dual_mov_b32 v115, s82
+; GFX11-NEXT:    v_dual_mov_b32 v114, s80 :: v_dual_mov_b32 v113, s81
+; GFX11-NEXT:    v_mov_b32_e32 v47, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 14
+; GFX11-NEXT:    v_dual_mov_b32 v23, s35 :: v_dual_mov_b32 v112, s71
+; GFX11-NEXT:    v_dual_mov_b32 v103, s70 :: v_dual_mov_b32 v102, s69
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v101, s67 :: v_dual_mov_b32 v46, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 15
+; GFX11-NEXT:    v_dual_mov_b32 v19, s34 :: v_dual_mov_b32 v100, s68
+; GFX11-NEXT:    v_dual_mov_b32 v99, s66 :: v_dual_mov_b32 v98, s65
+; GFX11-NEXT:    v_dual_mov_b32 v97, s64 :: v_dual_mov_b32 v44, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 16
+; GFX11-NEXT:    v_dual_mov_b32 v96, s54 :: v_dual_mov_b32 v87, s55
+; GFX11-NEXT:    v_dual_mov_b32 v86, s53 :: v_dual_mov_b32 v85, s52
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v45, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 17
+; GFX11-NEXT:    v_dual_mov_b32 v84, s51 :: v_dual_mov_b32 v83, s49
+; GFX11-NEXT:    v_dual_mov_b32 v147, s43 :: v_dual_mov_b32 v22, s78
+; GFX11-NEXT:    v_mov_b32_e32 v43, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 18
+; GFX11-NEXT:    v_dual_mov_b32 v67, s58 :: v_dual_mov_b32 v26, s88
+; GFX11-NEXT:    v_dual_mov_b32 v81, s44 :: v_dual_mov_b32 v30, s90
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_mov_b32_e32 v42, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 19
+; GFX11-NEXT:    v_dual_mov_b32 v17, s74 :: v_dual_mov_b32 v34, s92
+; GFX11-NEXT:    v_dual_mov_b32 v65, s94 :: v_dual_mov_b32 v68, s30
+; GFX11-NEXT:    v_mov_b32_e32 v41, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 20
+; GFX11-NEXT:    v_mov_b32_e32 v48, s62
+; GFX11-NEXT:    v_mov_b32_e32 v54, s72
+; GFX11-NEXT:    v_mov_b32_e32 v64, s60
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v70, s56 :: v_dual_mov_b32 v183, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 21
+; GFX11-NEXT:    v_mov_b32_e32 v80, s46
+; GFX11-NEXT:    v_mov_b32_e32 v18, s76
+; GFX11-NEXT:    v_mov_b32_e32 v40, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 22
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v182, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 23
+; GFX11-NEXT:    v_mov_b32_e32 v181, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 24
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v180, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 25
+; GFX11-NEXT:    v_mov_b32_e32 v178, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 26
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v179, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 27
+; GFX11-NEXT:    v_mov_b32_e32 v177, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 28
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v176, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 29
+; GFX11-NEXT:    v_mov_b32_e32 v167, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 30
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v165, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 31
+; GFX11-NEXT:    v_mov_b32_e32 v166, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v164, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 1
+; GFX11-NEXT:    v_mov_b32_e32 v163, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v162, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 3
+; GFX11-NEXT:    v_mov_b32_e32 v160, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v161, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 5
+; GFX11-NEXT:    v_mov_b32_e32 v151, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v150, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 7
+; GFX11-NEXT:    v_mov_b32_e32 v149, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v77, 8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_mov_b32_e32 v148, s0
+; GFX11-NEXT:    v_readlane_b32 s0, v78, 0
+; GFX11-NEXT:    v_readlane_b32 s1, v78, 1
+; GFX11-NEXT:    v_mov_b32_e32 v82, s0
+; GFX11-NEXT:  .LBB99_5: ; %end
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v74
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xff, v52
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 8, v82
+; GFX11-NEXT:    v_and_b32_e32 v53, 0xff, v53
+; GFX11-NEXT:    v_lshlrev_b32_e32 v82, 8, v63
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v81
+; GFX11-NEXT:    v_or_b32_e32 v52, v52, v69
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xff, v73
+; GFX11-NEXT:    v_and_b32_e32 v50, 0xff, v50
+; GFX11-NEXT:    v_and_b32_e32 v57, 0xff, v57
+; GFX11-NEXT:    v_lshlrev_b32_e32 v58, 8, v58
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v52
+; GFX11-NEXT:    v_or_b32_e32 v66, v69, v66
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v72
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v80
+; GFX11-NEXT:    v_and_b32_e32 v39, 0xff, v39
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
+; GFX11-NEXT:    v_or_b32_e32 v53, v53, v69
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xff, v62
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
+; GFX11-NEXT:    v_lshlrev_b32_e32 v67, 8, v67
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v32
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xff, v33
+; GFX11-NEXT:    v_or_b32_e32 v69, v69, v82
+; GFX11-NEXT:    v_lshlrev_b32_e32 v82, 8, v60
+; GFX11-NEXT:    v_and_b32_e32 v60, 0xff, v61
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 8, v64
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-NEXT:    v_or_b32_e32 v82, v50, v82
+; GFX11-NEXT:    v_or_b32_e32 v81, v60, v81
+; GFX11-NEXT:    v_or_b32_e32 v50, v52, v66
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v53
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 16, v69
+; GFX11-NEXT:    v_and_b32_e32 v66, 0xff, v51
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v59
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v25
+; GFX11-NEXT:    v_or_b32_e32 v51, v52, v53
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v82
+; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 16, v81
+; GFX11-NEXT:    v_or_b32_e32 v66, v66, v69
+; GFX11-NEXT:    v_or_b32_e32 v69, v57, v58
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v56
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xff, v47
+; GFX11-NEXT:    v_or_b32_e32 v52, v52, v53
+; GFX11-NEXT:    v_and_b32_e32 v53, 0xffff, v66
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v69
+; GFX11-NEXT:    v_or_b32_e32 v38, v38, v81
+; GFX11-NEXT:    v_or_b32_e32 v69, v82, v80
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v46
+; GFX11-NEXT:    v_and_b32_e32 v81, 0xff, v44
+; GFX11-NEXT:    v_lshlrev_b32_e32 v82, 8, v45
+; GFX11-NEXT:    v_or_b32_e32 v53, v53, v66
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v69
+; GFX11-NEXT:    v_or_b32_e32 v39, v39, v80
+; GFX11-NEXT:    v_or_b32_e32 v69, v81, v82
+; GFX11-NEXT:    v_and_b32_e32 v80, 0xff, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v43
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xff, v42
+; GFX11-NEXT:    v_or_b32_e32 v36, v38, v66
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v39
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v69
+; GFX11-NEXT:    v_or_b32_e32 v66, v80, v81
+; GFX11-NEXT:    v_or_b32_e32 v69, v82, v70
+; GFX11-NEXT:    v_and_b32_e32 v70, 0xff, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v80, 8, v41
+; GFX11-NEXT:    v_and_b32_e32 v81, 0xff, v183
+; GFX11-NEXT:    v_lshlrev_b32_e32 v82, 8, v40
+; GFX11-NEXT:    v_or_b32_e32 v37, v38, v39
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v66
+; GFX11-NEXT:    v_lshlrev_b32_e32 v39, 16, v69
+; GFX11-NEXT:    v_or_b32_e32 v66, v70, v80
+; GFX11-NEXT:    v_or_b32_e32 v69, v81, v82
+; GFX11-NEXT:    v_and_b32_e32 v80, 0xff, v181
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v182
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v180
+; GFX11-NEXT:    v_and_b32_e32 v66, 0xffff, v66
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
+; GFX11-NEXT:    v_or_b32_e32 v67, v80, v67
+; GFX11-NEXT:    v_or_b32_e32 v32, v32, v70
+; GFX11-NEXT:    v_or_b32_e32 v33, v33, v81
+; GFX11-NEXT:    v_or_b32_e32 v38, v38, v39
+; GFX11-NEXT:    v_or_b32_e32 v39, v66, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v67
+; GFX11-NEXT:    v_and_b32_e32 v67, 0xff, v178
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 8, v179
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v177
+; GFX11-NEXT:    v_and_b32_e32 v80, 0xff, v176
+; GFX11-NEXT:    v_lshlrev_b32_e32 v81, 8, v167
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xff, v165
+; GFX11-NEXT:    v_lshlrev_b32_e32 v165, 8, v166
+; GFX11-NEXT:    v_or_b32_e32 v67, v67, v69
+; GFX11-NEXT:    v_or_b32_e32 v28, v28, v70
+; GFX11-NEXT:    v_or_b32_e32 v64, v80, v64
+; GFX11-NEXT:    v_or_b32_e32 v29, v29, v81
+; GFX11-NEXT:    v_or_b32_e32 v69, v82, v165
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-NEXT:    v_lshlrev_b32_e32 v67, 16, v67
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    scratch_store_b128 v0, v[50:53], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[36:39], off offset:16
+; GFX11-NEXT:    v_or_b32_e32 v36, v32, v66
+; GFX11-NEXT:    v_or_b32_e32 v37, v33, v67
+; GFX11-NEXT:    v_or_b32_e32 v38, v28, v64
+; GFX11-NEXT:    v_or_b32_e32 v39, v29, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v164
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v163
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v54
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v162
+; GFX11-NEXT:    v_and_b32_e32 v50, 0xff, v160
+; GFX11-NEXT:    v_lshlrev_b32_e32 v51, 8, v161
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 8, v151
+; GFX11-NEXT:    v_or_b32_e32 v24, v24, v28
+; GFX11-NEXT:    v_or_b32_e32 v28, v29, v32
+; GFX11-NEXT:    v_or_b32_e32 v25, v25, v33
+; GFX11-NEXT:    v_or_b32_e32 v29, v50, v51
+; GFX11-NEXT:    v_or_b32_e32 v20, v20, v52
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v150
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v48
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v149
+; GFX11-NEXT:    v_and_b32_e32 v50, 0xff, v148
+; GFX11-NEXT:    v_lshlrev_b32_e32 v51, 8, v147
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v52, 8, v146
+; GFX11-NEXT:    v_and_b32_e32 v53, 0xff, v145
+; GFX11-NEXT:    v_lshlrev_b32_e32 v54, 8, v68
+; GFX11-NEXT:    v_or_b32_e32 v32, v32, v33
+; GFX11-NEXT:    v_or_b32_e32 v21, v21, v48
+; GFX11-NEXT:    v_or_b32_e32 v33, v50, v51
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v52
+; GFX11-NEXT:    v_or_b32_e32 v48, v53, v54
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
+; GFX11-NEXT:    v_or_b32_e32 v50, v24, v28
+; GFX11-NEXT:    v_or_b32_e32 v52, v20, v32
+; GFX11-NEXT:    v_or_b32_e32 v53, v21, v33
+; GFX11-NEXT:    v_or_b32_e32 v64, v15, v48
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v144
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v135
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v134
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 8, v133
+; GFX11-NEXT:    v_or_b32_e32 v51, v25, v29
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v132
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v65
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 8, v131
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v16, v20, v21
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v24
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v130
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 8, v129
+; GFX11-NEXT:    v_or_b32_e32 v20, v25, v28
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v29
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 8, v128
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v119
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 8, v34
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v118
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xff, v117
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 8, v116
+; GFX11-NEXT:    v_or_b32_e32 v21, v21, v24
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v25
+; GFX11-NEXT:    v_or_b32_e32 v24, v28, v29
+; GFX11-NEXT:    v_or_b32_e32 v12, v12, v32
+; GFX11-NEXT:    v_or_b32_e32 v25, v33, v34
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
+; GFX11-NEXT:    v_or_b32_e32 v65, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v66, v13, v20
+; GFX11-NEXT:    v_or_b32_e32 v67, v14, v21
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v115
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v114
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v30
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v113
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v112
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v103
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v24
+; GFX11-NEXT:    v_or_b32_e32 v12, v12, v25
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 8, v102
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v13
+; GFX11-NEXT:    v_or_b32_e32 v13, v14, v15
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v16
+; GFX11-NEXT:    v_or_b32_e32 v14, v20, v21
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v101
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v26
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v100
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 8, v97
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xff, v96
+; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 8, v22
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v24
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v99
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 8, v98
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v16
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v20
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v25
+; GFX11-NEXT:    v_or_b32_e32 v20, v26, v22
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_or_b32_e32 v16, v21, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_or_b32_e32 v13, v9, v13
+; GFX11-NEXT:    v_or_b32_e32 v14, v10, v14
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v15
+; GFX11-NEXT:    v_or_b32_e32 v9, v5, v20
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 8, v87
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v86
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v85
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v83
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v16
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v84
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v71
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v10, v15
+; GFX11-NEXT:    v_or_b32_e32 v10, v20, v18
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v16
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v49
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v55
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v35
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v31
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v27
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xff, v23
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v16
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v18
+; GFX11-NEXT:    v_or_b32_e32 v16, v20, v17
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v21
+; GFX11-NEXT:    v_or_b32_e32 v17, v22, v19
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-NEXT:    v_or_b32_e32 v10, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v1, v3, v15
+; GFX11-NEXT:    v_or_b32_e32 v2, v4, v18
+; GFX11-NEXT:    v_or_b32_e32 v3, v19, v16
+; GFX11-NEXT:    v_or_b32_e32 v4, v20, v17
+; GFX11-NEXT:    s_clause 0x5
+; GFX11-NEXT:    scratch_store_b128 v0, v[36:39], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[50:53], off offset:48
+; GFX11-NEXT:    scratch_store_b128 v0, v[64:67], off offset:64
+; GFX11-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
+; GFX11-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
+; GFX11-NEXT:    s_clause 0x12
+; GFX11-NEXT:    scratch_load_b32 v74, off, s32
+; GFX11-NEXT:    scratch_load_b32 v73, off, s32 offset:4
+; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:8
+; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:12
+; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:16
+; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:20
+; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:24
+; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:28
+; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:32
+; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:36
+; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:40
+; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:44
+; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:48
+; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:52
+; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:56
+; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:60
+; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:64
+; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:68
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:72
+; GFX11-NEXT:    v_readlane_b32 s104, v76, 8
+; GFX11-NEXT:    v_readlane_b32 s103, v76, 7
+; GFX11-NEXT:    v_readlane_b32 s102, v76, 6
+; GFX11-NEXT:    v_readlane_b32 s101, v76, 5
+; GFX11-NEXT:    v_readlane_b32 s100, v76, 4
+; GFX11-NEXT:    v_readlane_b32 s99, v76, 3
+; GFX11-NEXT:    v_readlane_b32 s98, v76, 2
+; GFX11-NEXT:    v_readlane_b32 s97, v76, 1
+; GFX11-NEXT:    v_readlane_b32 s96, v76, 0
+; GFX11-NEXT:    v_readlane_b32 s87, v75, 31
+; GFX11-NEXT:    v_readlane_b32 s86, v75, 30
+; GFX11-NEXT:    v_readlane_b32 s85, v75, 29
+; GFX11-NEXT:    v_readlane_b32 s84, v75, 28
+; GFX11-NEXT:    v_readlane_b32 s83, v75, 27
+; GFX11-NEXT:    v_readlane_b32 s82, v75, 26
+; GFX11-NEXT:    v_readlane_b32 s81, v75, 25
+; GFX11-NEXT:    v_readlane_b32 s80, v75, 24
+; GFX11-NEXT:    v_readlane_b32 s71, v75, 23
+; GFX11-NEXT:    v_readlane_b32 s70, v75, 22
+; GFX11-NEXT:    v_readlane_b32 s69, v75, 21
+; GFX11-NEXT:    v_readlane_b32 s68, v75, 20
+; GFX11-NEXT:    v_readlane_b32 s67, v75, 19
+; GFX11-NEXT:    v_readlane_b32 s66, v75, 18
+; GFX11-NEXT:    v_readlane_b32 s65, v75, 17
+; GFX11-NEXT:    v_readlane_b32 s64, v75, 16
+; GFX11-NEXT:    v_readlane_b32 s55, v75, 15
+; GFX11-NEXT:    v_readlane_b32 s54, v75, 14
+; GFX11-NEXT:    v_readlane_b32 s53, v75, 13
+; GFX11-NEXT:    v_readlane_b32 s52, v75, 12
+; GFX11-NEXT:    v_readlane_b32 s51, v75, 11
+; GFX11-NEXT:    v_readlane_b32 s50, v75, 10
+; GFX11-NEXT:    v_readlane_b32 s49, v75, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v75, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v75, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v75, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v75, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v75, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v75, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v75, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v75, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v75, 0
+; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:76
+; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:80
+; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:84
+; GFX11-NEXT:    scratch_load_b32 v78, off, s32 offset:88
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -231826,1395 +220282,700 @@ define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v64f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, v13 :: v_dual_mov_b32 v30, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB101_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB101_4
-; GFX11-TRUE16-NEXT:  .LBB101_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v18
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s12, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v1, 0x40c00000, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v3, 16, v17
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_cndmask_b32 v0, v5, v8
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v4, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v6, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v11, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v32
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v2
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v4, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v19
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v2, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v6 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v8, v3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v2, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v5 :: v_dual_and_b32 v6, 0xffff0000, v22
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v3, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v4 :: v_dual_and_b32 v35, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v23
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v2 :: v_dual_add_nc_u32 v2, v4, v5
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v2 :: v_dual_and_b32 v36, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v2, v4 :: v_dual_add_nc_u32 v0, v1, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v7, v2
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v6 :: v_dual_and_b32 v7, 0xffff0000, v25
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v5 :: v_dual_add_nc_u32 v5, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v2 :: v_dual_lshlrev_b32 v2, 16, v25
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v2
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v4, v5, v7 :: v_dual_add_nc_u32 v1, v3, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v6
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_nc_u32 v3, v3, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v3, v6 :: v_dual_and_b32 v5, 0xffff0000, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v7, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v2, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v4, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v3, 16, v28
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_cndmask_b32 v0, v0, v4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v0
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v4, v2
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v29
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v4, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v8, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v2, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v30
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v7
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v31
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v6 :: v_dual_add_f32 v1, 0x40c00000, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v31
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v4, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v4
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v3, v5 :: v_dual_add_f32 v3, 0x40c00000, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v7, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s12, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v4, v5, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s13, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v4, v6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s13, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v5
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s14, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v2, v6 :: v_dual_add_nc_u32 v5, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v4
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s14, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s15, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s15, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v9
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v10, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v5, v5, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s16, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v9, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s17, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v66
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v10, v9
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v4, v4, v7 :: v_dual_add_nc_u32 v7, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v67
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v54
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v6, v7, v8 :: v_dual_add_nc_u32 v7, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v11, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s18, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v7, v7, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v11, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v11
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v5, v8, v9 :: v_dual_add_nc_u32 v8, v10, v11
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s19, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v9, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s20, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v8, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, v13, v10
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v11, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v9, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v11
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s20, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v11
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s21, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v8, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v12, v13
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s21, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v11, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s22, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v10, v14, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, v15, v12
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v10
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v11, v14, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v10, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s22, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v11
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s23, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v10, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v14, v9
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v11, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s23, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, v13, v11
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v82, v14, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v12, v15, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v82, v82, v14
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v83, v12, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s24, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v68, 16, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v14
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v82, v83, v12
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v83, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s24, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v13, v13, v15 :: v_dual_add_nc_u32 v14, 0x7fff, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v82, v83, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v13
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s25, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v83
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v14, v15, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v82, v83
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v82, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s25, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, v15, v13
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v86, v82, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v83, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s26, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v14, v14, v85 :: v_dual_add_nc_u32 v15, 0x7fff, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v86, v86, v82
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v14
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v83, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v100, 0x400000, v83
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v15, v85, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v86
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v86, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s26, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v82
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v82, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s27, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v102, 0x400000, v86
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v96, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s27, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v97, v82, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v15, v15, v85, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v85, v86, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v99, v96, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v98, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v97, v97, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v103, 0x400000, v82
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v99, v99, v96
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v85, v85, v86
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v97, 0x7fff, v97
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v101, v98, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v112, 0x400000, v96
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v99, 0x7fff, v99
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v82, v97, v103 :: v_dual_add_nc_u32 v85, 0x7fff, v85
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v96, v96
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v101, v101, v98
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v14, v83
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v82
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v96, v99, v112, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v97, 0x7fff, v101
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v101, 0x400000, v98
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v96
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v85, v85, v102, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xffff, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v67
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v86, v97, v101 :: v_dual_and_b32 v65, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v85
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v80, 0xffff, v80
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v86
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v100, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xffff, v96
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v1, 16, v68
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v83, 16, v82
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v85, 16, v86
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v83, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xffff, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v96, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v85, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v87, 16, v83
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v9, 16, v86
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v97, 16, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v81, 16, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v81, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v83, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v84, 16, v85
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v69, 16, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v64, 16, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v70, 16, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v55, 16, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v50, 16, v65
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v52, 16, v55
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v38, 16, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v71, 16, v80
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v83
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v70
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v53, 16, v54
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v66
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v49
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v52
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v34, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v17, 16, v38
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB101_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB101_2
-; GFX11-TRUE16-NEXT:  .LBB101_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v64f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v31, v13 :: v_dual_mov_b32 v30, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s15, s3
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, s2
-; GFX11-FAKE16-NEXT:    s_mov_b32 s13, s1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s12, s0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB101_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB101_4
-; GFX11-FAKE16-NEXT:  .LBB101_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v18
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s12, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v1, 0x40c00000, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v7, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v6, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v3, 16, v17
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v5, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_cndmask_b32 v0, v5, v8
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v3, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v4, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v6, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v7, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v19
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v11, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v32
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v2
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v4, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v20
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v19
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v2, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v6 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v8, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v2, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v5 :: v_dual_and_b32 v6, 0xffff0000, v22
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v8, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v3, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v4 :: v_dual_add_nc_u32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v0, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v23
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v2 :: v_dual_add_nc_u32 v2, v4, v5
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v0, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v2 :: v_dual_and_b32 v36, 0xffff, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v2, v4 :: v_dual_add_nc_u32 v0, v1, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v3, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v2
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v6 :: v_dual_and_b32 v7, 0xffff0000, v25
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v6, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v5 :: v_dual_add_nc_u32 v5, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v7
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v2 :: v_dual_lshlrev_b32 v2, 16, v25
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v2
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v5, v7 :: v_dual_add_nc_u32 v1, v3, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v26
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_nc_u32 v3, v3, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xffff, v39
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v3, v6 :: v_dual_and_b32 v5, 0xffff0000, v27
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v27
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v7, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v2, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v4, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v28
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v3, 16, v28
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v49, 0xffff, v49
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_cndmask_b32 v0, v0, v4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v3, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v0
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v4, v2
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v29
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v4, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v8, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v2, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v30
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v30
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v7
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v31
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v0, v6 :: v_dual_add_f32 v1, 0x40c00000, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v31
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v0, v4, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v0, v4
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v3, v5 :: v_dual_add_f32 v3, 0x40c00000, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v7, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s12, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v4, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s13, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v4, v6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s13, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v5
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v7, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s14, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v6 :: v_dual_add_nc_u32 v5, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s14, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s15, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v8, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s15, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v5, v9
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v5, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v8, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s16, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v6
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v9, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s17, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xffff, v66
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v6
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v9
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v4, v7 :: v_dual_add_nc_u32 v7, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v67, 0xffff, v67
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v54, 0xffff, v54
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v7, v8 :: v_dual_add_nc_u32 v7, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v11, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s18, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v7, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v11, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v7
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v8, v9 :: v_dual_add_nc_u32 v8, v10, v11
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s19, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v10, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v9, v7
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v13, v10
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v11, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v9, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v11
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s20, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v11
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v13, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s21, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v12, v13
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s21, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, v11, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s22, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v14, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v12
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v10
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v11, v14, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s22, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v11
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s23, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v14, v9
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v11, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s23, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v13, v11
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v82, v14, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v12, v15, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v11
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v82, v82, v14
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v83, v12, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v68, 16, v5
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v82
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v14
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v82, v83, v12
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v83, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s24, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v13, v13, v15 :: v_dual_add_nc_u32 v14, 0x7fff, v82
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v82, v83, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v13
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s25, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v85, 0x400000, v83
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v14, v15, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v82, v83
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v13, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v82, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s25, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v13
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v86, v82, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v83, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s26, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v85 :: v_dual_add_nc_u32 v15, 0x7fff, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v85, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v86, v86, v82
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v14
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v83, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v100, 0x400000, v83
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v15, v85, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v86
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v86, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s26, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v85, 0x400000, v82
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v82, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s27, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v102, 0x400000, v86
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v96, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s27, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v97, v82, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v85, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v85, v86, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v99, v96, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v98, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, v97, v82
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v103, 0x400000, v82
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v99, v99, v96
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v85, v85, v86
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, 0x7fff, v97
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v101, v98, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v112, 0x400000, v96
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v99, 0x7fff, v99
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v82, v97, v103 :: v_dual_add_nc_u32 v85, 0x7fff, v85
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v96, v96
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v101, v101, v98
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v14, v83
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v82
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v96, v99, v112, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, 0x7fff, v101
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v101, 0x400000, v98
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v96
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v85, v85, v102, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xffff, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v68, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v67
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v86, v97, v101 :: v_dual_and_b32 v65, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v85
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v80, 0xffff, v80
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v86
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v100, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v86, 0xffff, v96
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v1, 16, v68
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v83, 16, v82
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v85, 16, v86
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v83, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v86, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xffff, v96
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v96, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v85, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v87, 16, v83
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v9, 16, v86
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v97, 16, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v81, 16, v96
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v81, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v83, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v84, 16, v85
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v69, 16, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v64, 16, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v70, 16, v81
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v70, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v55, 16, v69
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v55, 0xffff, v30
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v64
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v50, 16, v65
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v52, 16, v55
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v38, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v51
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v71, 16, v80
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v83
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v70
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v53, 16, v54
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v66
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v49
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v50
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v52
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v17, 16, v38
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB101_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB101_2
-; GFX11-FAKE16-NEXT:  .LBB101_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v64bf16_to_v64f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    v_dual_mov_b32 v31, v13 :: v_dual_mov_b32 v30, v12
+; GFX11-NEXT:    v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v28, v10
+; GFX11-NEXT:    v_dual_mov_b32 v27, v9 :: v_dual_mov_b32 v26, v8
+; GFX11-NEXT:    v_dual_mov_b32 v25, v7 :: v_dual_mov_b32 v24, v6
+; GFX11-NEXT:    v_dual_mov_b32 v23, v5 :: v_dual_mov_b32 v22, v4
+; GFX11-NEXT:    v_dual_mov_b32 v21, v3 :: v_dual_mov_b32 v20, v2
+; GFX11-NEXT:    v_dual_mov_b32 v19, v1 :: v_dual_mov_b32 v18, v0
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s15, s3
+; GFX11-NEXT:    s_mov_b32 s14, s2
+; GFX11-NEXT:    s_mov_b32 s13, s1
+; GFX11-NEXT:    s_mov_b32 s12, s0
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB101_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB101_4
+; GFX11-NEXT:  .LBB101_2: ; %cmp.true
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff0000, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff0000, v18
+; GFX11-NEXT:    s_and_b32 s0, s12, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v1, 0x40c00000, v1
+; GFX11-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v6, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, v7, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, v6, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff0000, v16
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v3, 16, v17
+; GFX11-NEXT:    v_bfe_u32 v5, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v5, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_cndmask_b32 v0, v5, v8
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_bfe_u32 v11, v3, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v5, v4, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v6, v9, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v7, v10, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 16, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v18
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v5, v4
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff0000, v19
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, v11, v3
+; GFX11-NEXT:    v_lshl_or_b32 v16, v16, 16, v32
+; GFX11-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_bfe_u32 v3, v2, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_bfe_u32 v4, v5, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v2
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v4, v5
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff0000, v20
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v19
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_bfe_u32 v2, v4, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 16, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v2, v4
+; GFX11-NEXT:    v_dual_add_f32 v2, 0x40c00000, v6 :: v_dual_add_f32 v3, 0x40c00000, v3
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff0000, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_bfe_u32 v8, v3, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v8, v3
+; GFX11-NEXT:    v_bfe_u32 v8, v2, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v3
+; GFX11-NEXT:    v_add_f32_e32 v3, 0x40c00000, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v1, v5 :: v_dual_and_b32 v6, 0xffff0000, v22
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v8, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v7, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v1, v3, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v3
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v3
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v0, v4 :: v_dual_add_nc_u32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_bfe_u32 v4, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_add_f32_e32 v3, 0x40c00000, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v0
+; GFX11-NEXT:    v_bfe_u32 v0, v6, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff0000, v23
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v1, v2 :: v_dual_add_nc_u32 v2, v4, v5
+; GFX11-NEXT:    v_bfe_u32 v4, v3, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, v0, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v2
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
+; GFX11-NEXT:    v_add_f32_e32 v5, 0x40c00000, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v1, v2 :: v_dual_and_b32 v36, 0xffff, v36
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v19, v19, 16, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_bfe_u32 v1, v5, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff0000, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v2, v4 :: v_dual_add_nc_u32 v0, v1, v5
+; GFX11-NEXT:    v_add_f32_e32 v2, 0x40c00000, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v5
+; GFX11-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
+; GFX11-NEXT:    v_bfe_u32 v7, v2, 16, 1
+; GFX11-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v1
+; GFX11-NEXT:    v_bfe_u32 v1, v3, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v7, v2
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v0, v6 :: v_dual_and_b32 v7, 0xffff0000, v25
+; GFX11-NEXT:    v_bfe_u32 v6, v4, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, v6, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v0, v5 :: v_dual_add_nc_u32 v5, 0x7fff, v6
+; GFX11-NEXT:    v_add_f32_e32 v6, 0x40c00000, v7
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v4
+; GFX11-NEXT:    v_bfe_u32 v3, v6, 16, 1
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v1, v2 :: v_dual_lshlrev_b32 v2, 16, v25
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v1
+; GFX11-NEXT:    v_add_f32_e32 v0, 0x40c00000, v2
+; GFX11-NEXT:    v_dual_cndmask_b32 v4, v5, v7 :: v_dual_add_nc_u32 v1, v3, v6
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff0000, v26
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v3, v0, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v26
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_nc_u32 v3, v3, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v0
+; GFX11-NEXT:    v_and_b32_e32 v39, 0xffff, v39
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v7, v4, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v5, v2
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v3, v6 :: v_dual_and_b32 v5, 0xffff0000, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v27
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v7, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v2
+; GFX11-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
+; GFX11-NEXT:    v_bfe_u32 v2, v5, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_bfe_u32 v4, v6, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, v2, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v1
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v8, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v4, v6
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff0000, v28
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v3, 16, v28
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_and_b32_e32 v49, 0xffff, v49
+; GFX11-NEXT:    v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_cndmask_b32 v0, v0, v4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    v_bfe_u32 v4, v2, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v5, v3, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v0
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v4, v2
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff0000, v29
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v5, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v29
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
+; GFX11-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v5, 0x40c00000, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v8, v4, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_bfe_u32 v2, v5, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v8, v4
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, v2, v5
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff0000, v30
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
+; GFX11-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v30
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
+; GFX11-NEXT:    v_or_b32_e32 v3, 0x400000, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_add_f32_e32 v4, 0x40c00000, v7
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v3, v2, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff0000, v31
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v0, v6 :: v_dual_add_f32 v1, 0x40c00000, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v31
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v0
+; GFX11-NEXT:    v_bfe_u32 v0, v4, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_bfe_u32 v7, v1, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, v0, v4
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v3, v5 :: v_dual_add_f32 v3, 0x40c00000, v6
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, v7, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
+; GFX11-NEXT:    v_bfe_u32 v2, v3, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v6
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s12, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v0
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v4, v5, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v4, v6, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s13, 0xffff0000
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v4, v6
+; GFX11-NEXT:    v_bfe_u32 v2, v5, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    s_lshl_b32 s0, s13, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v5
+; GFX11-NEXT:    v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, v7, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v1
+; GFX11-NEXT:    v_bfe_u32 v1, v4, 16, 1
+; GFX11-NEXT:    s_and_b32 s0, s14, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v2, v6 :: v_dual_add_nc_u32 v5, 0x7fff, v7
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v4
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    s_lshl_b32 s0, s14, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v8, v7, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v1
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    s_and_b32 s0, s15, 0xffff0000
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v8, v7
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v7
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v5, v9, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    s_lshl_b32 s0, s15, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v5, v9
+; GFX11-NEXT:    v_bfe_u32 v10, v6, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, v10, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v4
+; GFX11-NEXT:    v_bfe_u32 v4, v7, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v6
+; GFX11-NEXT:    s_and_b32 s0, s16, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v5, v5, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v10
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v5
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    v_cndmask_b32_e32 v6, v8, v9, vcc_lo
+; GFX11-NEXT:    s_lshl_b32 s0, s16, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v7
+; GFX11-NEXT:    v_bfe_u32 v9, v5, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v6
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    s_and_b32 s0, s17, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, v9, v5
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s17, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v8, v6, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_bfe_u32 v10, v9, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v66, 0xffff, v66
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v7
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v6
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, v10, v9
+; GFX11-NEXT:    s_and_b32 s0, s18, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v4, v4, v7 :: v_dual_add_nc_u32 v7, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    v_bfe_u32 v11, v5, 16, 1
+; GFX11-NEXT:    v_and_b32_e32 v67, 0xffff, v67
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v54, 0xffff, v54
+; GFX11-NEXT:    v_dual_cndmask_b32 v6, v7, v8 :: v_dual_add_nc_u32 v7, 0x7fff, v10
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v9
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, v11, v5
+; GFX11-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s18, 16
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v7, v7, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v10
+; GFX11-NEXT:    v_bfe_u32 v10, v11, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v7
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s19, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v5, v8, v9 :: v_dual_add_nc_u32 v8, v10, v11
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v9, v7, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-NEXT:    s_lshl_b32 s0, s19, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
+; GFX11-NEXT:    v_bfe_u32 v13, v10, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, v9, v7
+; GFX11-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s20, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v8, v12, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, v13, v10
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v8
+; GFX11-NEXT:    v_bfe_u32 v8, v11, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v9, v12, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v13
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v10
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v11
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s20, 16
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v11
+; GFX11-NEXT:    v_cndmask_b32_e32 v9, v9, v12, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
+; GFX11-NEXT:    v_bfe_u32 v12, v13, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v9
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s21, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v8, v10, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, v12, v13
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v11, v9, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    s_lshl_b32 s0, s21, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
+; GFX11-NEXT:    v_bfe_u32 v15, v12, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, v11, v9
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s22, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v10, v10, v14, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v15, v15, v12
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v10
+; GFX11-NEXT:    v_bfe_u32 v10, v13, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_cndmask_b32_e32 v9, v11, v14, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v15
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v12
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, v10, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v9
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s22, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v13
+; GFX11-NEXT:    v_bfe_u32 v14, v9, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 16, v11
+; GFX11-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s23, 0xffff0000
+; GFX11-NEXT:    v_or_b32_e32 v15, 0x400000, v9
+; GFX11-NEXT:    v_cndmask_b32_e32 v10, v10, v12, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, v14, v9
+; GFX11-NEXT:    v_bfe_u32 v13, v11, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    s_lshl_b32 s0, s23, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, v13, v11
+; GFX11-NEXT:    v_bfe_u32 v82, v14, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_cndmask_b32_e32 v9, v12, v15, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
+; GFX11-NEXT:    v_or_b32_e32 v15, 0x400000, v11
+; GFX11-NEXT:    v_add_nc_u32_e32 v82, v82, v14
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-NEXT:    v_bfe_u32 v83, v12, 16, 1
+; GFX11-NEXT:    s_and_b32 s0, s24, 0xffff0000
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v5, v68, 16, v5
+; GFX11-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v82
+; GFX11-NEXT:    v_or_b32_e32 v15, 0x400000, v14
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
+; GFX11-NEXT:    v_add_nc_u32_e32 v82, v83, v12
+; GFX11-NEXT:    v_add_f32_e64 v83, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s24, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v13, v13, v15 :: v_dual_add_nc_u32 v14, 0x7fff, v82
+; GFX11-NEXT:    v_or_b32_e32 v15, 0x400000, v12
+; GFX11-NEXT:    v_bfe_u32 v82, v83, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 16, v13
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s25, 0xffff0000
+; GFX11-NEXT:    v_or_b32_e32 v85, 0x400000, v83
+; GFX11-NEXT:    v_cndmask_b32_e32 v12, v14, v15, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v82, v83
+; GFX11-NEXT:    v_bfe_u32 v15, v13, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v82, 0x40c00000, s0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
+; GFX11-NEXT:    s_lshl_b32 s0, s25, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
+; GFX11-NEXT:    v_add_nc_u32_e32 v15, v15, v13
+; GFX11-NEXT:    v_bfe_u32 v86, v82, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v83, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s26, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v14, v14, v85 :: v_dual_add_nc_u32 v15, 0x7fff, v15
+; GFX11-NEXT:    v_or_b32_e32 v85, 0x400000, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v86, v86, v82
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 16, v14
+; GFX11-NEXT:    v_bfe_u32 v14, v83, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v100, 0x400000, v83
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v15, v85, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v86
+; GFX11-NEXT:    v_add_f32_e64 v86, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s26, 16
+; GFX11-NEXT:    v_or_b32_e32 v85, 0x400000, v82
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
+; GFX11-NEXT:    v_add_f32_e64 v82, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s27, 16
+; GFX11-NEXT:    v_or_b32_e32 v102, 0x400000, v86
+; GFX11-NEXT:    v_add_f32_e64 v96, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s27, 0xffff0000
+; GFX11-NEXT:    v_bfe_u32 v97, v82, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v15, v15, v85, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v85, v86, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v99, v96, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v98, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v97, v97, v82
+; GFX11-NEXT:    v_or_b32_e32 v103, 0x400000, v82
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
+; GFX11-NEXT:    v_add_nc_u32_e32 v99, v99, v96
+; GFX11-NEXT:    v_add_nc_u32_e32 v85, v85, v86
+; GFX11-NEXT:    v_add_nc_u32_e32 v97, 0x7fff, v97
+; GFX11-NEXT:    v_bfe_u32 v101, v98, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v112, 0x400000, v96
+; GFX11-NEXT:    v_add_nc_u32_e32 v99, 0x7fff, v99
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v82, v97, v103 :: v_dual_add_nc_u32 v85, 0x7fff, v85
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v96, v96
+; GFX11-NEXT:    v_add_nc_u32_e32 v101, v101, v98
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v14, v83
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v82
+; GFX11-NEXT:    v_cndmask_b32_e32 v96, v99, v112, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
+; GFX11-NEXT:    v_add_nc_u32_e32 v97, 0x7fff, v101
+; GFX11-NEXT:    v_or_b32_e32 v101, 0x400000, v98
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 16, v96
+; GFX11-NEXT:    v_cndmask_b32_e32 v85, v85, v102, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xffff, v82
+; GFX11-NEXT:    v_and_b32_e32 v68, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v3, v65, 16, v67
+; GFX11-NEXT:    v_dual_cndmask_b32 v86, v97, v101 :: v_dual_and_b32 v65, 0xffff, v28
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v85
+; GFX11-NEXT:    v_lshrrev_b32_e32 v97, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v80, 0xffff, v80
+; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 16, v86
+; GFX11-NEXT:    v_cndmask_b32_e32 v14, v14, v100, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v86, 0xffff, v96
+; GFX11-NEXT:    v_lshl_or_b32 v1, v1, 16, v68
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v14, v83, 16, v82
+; GFX11-NEXT:    v_lshl_or_b32 v15, v85, 16, v86
+; GFX11-NEXT:    v_and_b32_e32 v83, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v86, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xffff, v96
+; GFX11-NEXT:    v_and_b32_e32 v96, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v85, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v87, 16, v83
+; GFX11-NEXT:    v_lshl_or_b32 v10, v9, 16, v86
+; GFX11-NEXT:    v_lshl_or_b32 v13, v97, 16, v82
+; GFX11-NEXT:    v_and_b32_e32 v82, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v9, v81, 16, v96
+; GFX11-NEXT:    v_and_b32_e32 v81, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v83, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v11, v84, 16, v85
+; GFX11-NEXT:    v_lshl_or_b32 v6, v69, 16, v82
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v2, v64, 16, v66
+; GFX11-NEXT:    v_and_b32_e32 v64, 0xffff, v29
+; GFX11-NEXT:    v_lshl_or_b32 v7, v70, 16, v81
+; GFX11-NEXT:    v_and_b32_e32 v70, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v0, v55, 16, v69
+; GFX11-NEXT:    v_and_b32_e32 v55, 0xffff, v30
+; GFX11-NEXT:    v_lshl_or_b32 v28, v51, 16, v64
+; GFX11-NEXT:    v_and_b32_e32 v51, 0xffff, v24
+; GFX11-NEXT:    v_and_b32_e32 v66, 0xffff, v27
+; GFX11-NEXT:    v_lshl_or_b32 v27, v50, 16, v65
+; GFX11-NEXT:    v_lshl_or_b32 v29, v52, 16, v55
+; GFX11-NEXT:    v_and_b32_e32 v50, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v52, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v24, v38, 16, v39
+; GFX11-NEXT:    v_lshl_or_b32 v22, v37, 16, v51
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v38, 0xffff, v18
+; GFX11-NEXT:    v_lshl_or_b32 v8, v71, 16, v80
+; GFX11-NEXT:    v_lshl_or_b32 v4, v4, 16, v83
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v70
+; GFX11-NEXT:    v_lshl_or_b32 v30, v53, 16, v54
+; GFX11-NEXT:    v_lshl_or_b32 v26, v26, 16, v66
+; GFX11-NEXT:    v_lshl_or_b32 v25, v48, 16, v49
+; GFX11-NEXT:    v_lshl_or_b32 v23, v23, 16, v50
+; GFX11-NEXT:    v_lshl_or_b32 v21, v21, 16, v52
+; GFX11-NEXT:    v_lshl_or_b32 v20, v35, 16, v36
+; GFX11-NEXT:    v_lshl_or_b32 v18, v33, 16, v37
+; GFX11-NEXT:    v_lshl_or_b32 v17, v17, 16, v38
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB101_3:
+; GFX11-NEXT:    s_branch .LBB101_2
+; GFX11-NEXT:  .LBB101_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
+; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll
index 1056c66720b25..34d7ed9290b67 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll
@@ -3739,140 +3739,69 @@ define inreg <16 x i8> @bitcast_v4i32_to_v16i8_scalar(<4 x i32> inreg %a, i32 in
 ; GFX9-NEXT:    ; implicit-def: $sgpr10
 ; GFX9-NEXT:    s_branch .LBB25_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4i32_to_v16i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB25_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s18
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB25_3
-; GFX11-TRUE16-NEXT:  .LBB25_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB25_3: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s8
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB25_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr17_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr16_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB25_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v4i32_to_v16i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s18, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB25_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s18
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB25_3
-; GFX11-FAKE16-NEXT:  .LBB25_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB25_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s16 :: v_dual_mov_b32 v3, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s14 :: v_dual_mov_b32 v7, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s3 :: v_dual_mov_b32 v13, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s9 :: v_dual_mov_b32 v15, s8
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB25_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr17
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr16
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    s_branch .LBB25_2
+; GFX11-LABEL: bitcast_v4i32_to_v16i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s16, 0
+; GFX11-NEXT:    s_mov_b32 s18, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB25_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s8, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s9, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s11, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s14, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s16, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s17, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s18
+; GFX11-NEXT:    s_cbranch_vccnz .LBB25_3
+; GFX11-NEXT:  .LBB25_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s8, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s9, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s11, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s14, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s16, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s17, s0, 8
+; GFX11-NEXT:  .LBB25_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s17
+; GFX11-NEXT:    v_dual_mov_b32 v2, s16 :: v_dual_mov_b32 v3, s6
+; GFX11-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s15
+; GFX11-NEXT:    v_dual_mov_b32 v6, s14 :: v_dual_mov_b32 v7, s13
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s12
+; GFX11-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s4
+; GFX11-NEXT:    v_dual_mov_b32 v12, s3 :: v_dual_mov_b32 v13, s10
+; GFX11-NEXT:    v_dual_mov_b32 v14, s9 :: v_dual_mov_b32 v15, s8
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB25_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr17
+; GFX11-NEXT:    ; implicit-def: $sgpr16
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    s_branch .LBB25_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -8282,158 +8211,78 @@ define inreg <16 x i8> @bitcast_v4f32_to_v16i8_scalar(<4 x float> inreg %a, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v12, v17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4f32_to_v16i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s18, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, s1, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, s3, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, s2, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, s0, 1.0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v12
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_5
-; GFX11-TRUE16-NEXT:  .LBB49_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr18_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr17_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr16_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s4
-; GFX11-TRUE16-NEXT:  .LBB49_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v16.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4f32_to_v16i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s18, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, s1, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, s3, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, s2, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, s0, 1.0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_5
-; GFX11-FAKE16-NEXT:  .LBB49_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr18
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr16
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr17
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s0 :: v_dual_mov_b32 v19, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v17, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s18 :: v_dual_mov_b32 v2, s16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s14 :: v_dual_mov_b32 v10, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s13 :: v_dual_mov_b32 v14, s10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v13, s12
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v15, s9
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB49_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v18
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v19
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v16
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v17
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4f32_to_v16i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s16, 0
+; GFX11-NEXT:    s_mov_b32 s8, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s9, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s10, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s11, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s14, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s15, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s17, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s16, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s18, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_4
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f32_e64 v19, s1, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v17, s3, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v16, s2, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v18, s0, 1.0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v17
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[16:17]
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[18:19]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v18
+; GFX11-NEXT:    s_branch .LBB49_5
+; GFX11-NEXT:  .LBB49_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr18
+; GFX11-NEXT:    ; implicit-def: $sgpr16
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr17
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    s_branch .LBB49_2
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    v_dual_mov_b32 v18, s0 :: v_dual_mov_b32 v19, s1
+; GFX11-NEXT:    v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v17, s3
+; GFX11-NEXT:    v_dual_mov_b32 v1, s18 :: v_dual_mov_b32 v2, s16
+; GFX11-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s15
+; GFX11-NEXT:    v_dual_mov_b32 v7, s14 :: v_dual_mov_b32 v10, s11
+; GFX11-NEXT:    v_dual_mov_b32 v9, s13 :: v_dual_mov_b32 v14, s10
+; GFX11-NEXT:    v_mov_b32_e32 v13, s12
+; GFX11-NEXT:    v_mov_b32_e32 v15, s9
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB49_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v18
+; GFX11-NEXT:    v_mov_b32_e32 v4, v19
+; GFX11-NEXT:    v_mov_b32_e32 v8, v16
+; GFX11-NEXT:    v_mov_b32_e32 v12, v17
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -12468,140 +12317,69 @@ define inreg <16 x i8> @bitcast_v2i64_to_v16i8_scalar(<2 x i64> inreg %a, i32 in
 ; GFX9-NEXT:    ; implicit-def: $sgpr10
 ; GFX9-NEXT:    s_branch .LBB69_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v2i64_to_v16i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB69_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s18
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB69_3
-; GFX11-TRUE16-NEXT:  .LBB69_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB69_3: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s8
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB69_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr17_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr16_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB69_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v2i64_to_v16i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s18, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB69_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s18
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB69_3
-; GFX11-FAKE16-NEXT:  .LBB69_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB69_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s16 :: v_dual_mov_b32 v3, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s14 :: v_dual_mov_b32 v7, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s3 :: v_dual_mov_b32 v13, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s9 :: v_dual_mov_b32 v15, s8
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB69_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr17
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr16
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    s_branch .LBB69_2
+; GFX11-LABEL: bitcast_v2i64_to_v16i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s16, 0
+; GFX11-NEXT:    s_mov_b32 s18, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB69_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s8, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s9, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s11, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s14, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s16, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s17, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s18
+; GFX11-NEXT:    s_cbranch_vccnz .LBB69_3
+; GFX11-NEXT:  .LBB69_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b32 s8, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s9, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s11, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s14, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s16, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s17, s0, 8
+; GFX11-NEXT:  .LBB69_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s17
+; GFX11-NEXT:    v_dual_mov_b32 v2, s16 :: v_dual_mov_b32 v3, s6
+; GFX11-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s15
+; GFX11-NEXT:    v_dual_mov_b32 v6, s14 :: v_dual_mov_b32 v7, s13
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s12
+; GFX11-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s4
+; GFX11-NEXT:    v_dual_mov_b32 v12, s3 :: v_dual_mov_b32 v13, s10
+; GFX11-NEXT:    v_dual_mov_b32 v14, s9 :: v_dual_mov_b32 v15, s8
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB69_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr17
+; GFX11-NEXT:    ; implicit-def: $sgpr16
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    s_branch .LBB69_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -16240,151 +16018,75 @@ define inreg <16 x i8> @bitcast_v2f64_to_v16i8_scalar(<2 x double> inreg %a, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v12, v17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v2f64_to_v16i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB85_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s18, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB85_4
-; GFX11-TRUE16-NEXT:  .LBB85_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[11:12], s[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[3:4], s[0:1], 1.0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB85_5
-; GFX11-TRUE16-NEXT:  .LBB85_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr18_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr17_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr16_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB85_2
-; GFX11-TRUE16-NEXT:  .LBB85_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s14
-; GFX11-TRUE16-NEXT:  .LBB85_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v16.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v2f64_to_v16i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB85_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s18, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB85_4
-; GFX11-FAKE16-NEXT:  .LBB85_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], s[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], s[0:1], 1.0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-FAKE16-NEXT:    s_branch .LBB85_5
-; GFX11-FAKE16-NEXT:  .LBB85_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr17
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr18
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr16
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    s_branch .LBB85_2
-; GFX11-FAKE16-NEXT:  .LBB85_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s0 :: v_dual_mov_b32 v17, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v19, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s18 :: v_dual_mov_b32 v1, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s16 :: v_dual_mov_b32 v9, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_mov_b32 v14, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s6 :: v_dual_mov_b32 v6, s10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v15, s14
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v13, s12
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v7, s11
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v5, s9
-; GFX11-FAKE16-NEXT:  .LBB85_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v18
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v19
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v16
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v17
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v2f64_to_v16i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s16, 0
+; GFX11-NEXT:    s_mov_b32 s8, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB85_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s14, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s13, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s16, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s11, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s10, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s18, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s17, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
+; GFX11-NEXT:    s_cbranch_vccnz .LBB85_4
+; GFX11-NEXT:  .LBB85_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[16:17], s[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], s[0:1], 1.0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[16:17]
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[18:19]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v18
+; GFX11-NEXT:    s_branch .LBB85_5
+; GFX11-NEXT:  .LBB85_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr17
+; GFX11-NEXT:    ; implicit-def: $sgpr18
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr16
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    s_branch .LBB85_2
+; GFX11-NEXT:  .LBB85_4:
+; GFX11-NEXT:    v_dual_mov_b32 v18, s0 :: v_dual_mov_b32 v17, s3
+; GFX11-NEXT:    v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v19, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s18 :: v_dual_mov_b32 v1, s17
+; GFX11-NEXT:    v_dual_mov_b32 v10, s16 :: v_dual_mov_b32 v9, s15
+; GFX11-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_mov_b32 v14, s13
+; GFX11-NEXT:    v_dual_mov_b32 v11, s6 :: v_dual_mov_b32 v6, s10
+; GFX11-NEXT:    v_mov_b32_e32 v15, s14
+; GFX11-NEXT:    v_mov_b32_e32 v13, s12
+; GFX11-NEXT:    v_mov_b32_e32 v7, s11
+; GFX11-NEXT:    v_mov_b32_e32 v5, s9
+; GFX11-NEXT:  .LBB85_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v18
+; GFX11-NEXT:    v_mov_b32_e32 v4, v19
+; GFX11-NEXT:    v_mov_b32_e32 v8, v16
+; GFX11-NEXT:    v_mov_b32_e32 v12, v17
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -19733,158 +19435,78 @@ define inreg <16 x i8> @bitcast_v8i16_to_v16i8_scalar(<8 x i16> inreg %a, i32 in
 ; GFX9-NEXT:    v_mov_b32_e32 v12, v17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v8i16_to_v16i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB97_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s18, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB97_4
-; GFX11-TRUE16-NEXT:  .LBB97_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v12
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB97_5
-; GFX11-TRUE16-NEXT:  .LBB97_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr18_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr17_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr16_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB97_2
-; GFX11-TRUE16-NEXT:  .LBB97_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s4
-; GFX11-TRUE16-NEXT:  .LBB97_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v16.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v8i16_to_v16i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB97_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s18, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB97_4
-; GFX11-FAKE16-NEXT:  .LBB97_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v19, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v17, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v16, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v18, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-FAKE16-NEXT:    s_branch .LBB97_5
-; GFX11-FAKE16-NEXT:  .LBB97_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr18
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr16
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr17
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    s_branch .LBB97_2
-; GFX11-FAKE16-NEXT:  .LBB97_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s0 :: v_dual_mov_b32 v19, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v17, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s18 :: v_dual_mov_b32 v2, s16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s14 :: v_dual_mov_b32 v10, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s13 :: v_dual_mov_b32 v14, s10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v13, s12
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v15, s9
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB97_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v18
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v19
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v16
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v17
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v8i16_to_v16i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s16, 0
+; GFX11-NEXT:    s_mov_b32 s8, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB97_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s9, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s10, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s11, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s14, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s15, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s17, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s16, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s18, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
+; GFX11-NEXT:    s_cbranch_vccnz .LBB97_4
+; GFX11-NEXT:  .LBB97_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v19, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v17, s3, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v16, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v18, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v17
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[16:17]
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[18:19]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v18
+; GFX11-NEXT:    s_branch .LBB97_5
+; GFX11-NEXT:  .LBB97_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr18
+; GFX11-NEXT:    ; implicit-def: $sgpr16
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr17
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    s_branch .LBB97_2
+; GFX11-NEXT:  .LBB97_4:
+; GFX11-NEXT:    v_dual_mov_b32 v18, s0 :: v_dual_mov_b32 v19, s1
+; GFX11-NEXT:    v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v17, s3
+; GFX11-NEXT:    v_dual_mov_b32 v1, s18 :: v_dual_mov_b32 v2, s16
+; GFX11-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s15
+; GFX11-NEXT:    v_dual_mov_b32 v7, s14 :: v_dual_mov_b32 v10, s11
+; GFX11-NEXT:    v_dual_mov_b32 v9, s13 :: v_dual_mov_b32 v14, s10
+; GFX11-NEXT:    v_mov_b32_e32 v13, s12
+; GFX11-NEXT:    v_mov_b32_e32 v15, s9
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB97_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v18
+; GFX11-NEXT:    v_mov_b32_e32 v4, v19
+; GFX11-NEXT:    v_mov_b32_e32 v8, v16
+; GFX11-NEXT:    v_mov_b32_e32 v12, v17
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -22748,158 +22370,78 @@ define inreg <16 x i8> @bitcast_v8f16_to_v16i8_scalar(<8 x half> inreg %a, i32 i
 ; GFX9-NEXT:    v_mov_b32_e32 v12, v17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v8f16_to_v16i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB105_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s18, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB105_4
-; GFX11-TRUE16-NEXT:  .LBB105_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v12
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB105_5
-; GFX11-TRUE16-NEXT:  .LBB105_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr18_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr17_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr16_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB105_2
-; GFX11-TRUE16-NEXT:  .LBB105_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s4
-; GFX11-TRUE16-NEXT:  .LBB105_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v16.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v8f16_to_v16i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB105_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s18, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB105_4
-; GFX11-FAKE16-NEXT:  .LBB105_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v19, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v17, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v16, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v18, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v17
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[16:17]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[18:19]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v18
-; GFX11-FAKE16-NEXT:    s_branch .LBB105_5
-; GFX11-FAKE16-NEXT:  .LBB105_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr18
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr16
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr17
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    s_branch .LBB105_2
-; GFX11-FAKE16-NEXT:  .LBB105_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s0 :: v_dual_mov_b32 v19, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v17, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s18 :: v_dual_mov_b32 v2, s16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s14 :: v_dual_mov_b32 v10, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s13 :: v_dual_mov_b32 v14, s10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v13, s12
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v15, s9
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB105_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v18
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v19
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v16
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v17
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v8f16_to_v16i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s16, 0
+; GFX11-NEXT:    s_mov_b32 s8, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB105_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s9, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s10, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s11, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s14, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s15, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s17, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s16, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s18, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
+; GFX11-NEXT:    s_cbranch_vccnz .LBB105_4
+; GFX11-NEXT:  .LBB105_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v17
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[16:17]
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[18:19]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v18
+; GFX11-NEXT:    s_branch .LBB105_5
+; GFX11-NEXT:  .LBB105_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr18
+; GFX11-NEXT:    ; implicit-def: $sgpr16
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr17
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    s_branch .LBB105_2
+; GFX11-NEXT:  .LBB105_4:
+; GFX11-NEXT:    v_dual_mov_b32 v18, s0 :: v_dual_mov_b32 v19, s1
+; GFX11-NEXT:    v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v17, s3
+; GFX11-NEXT:    v_dual_mov_b32 v1, s18 :: v_dual_mov_b32 v2, s16
+; GFX11-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s15
+; GFX11-NEXT:    v_dual_mov_b32 v7, s14 :: v_dual_mov_b32 v10, s11
+; GFX11-NEXT:    v_dual_mov_b32 v9, s13 :: v_dual_mov_b32 v14, s10
+; GFX11-NEXT:    v_mov_b32_e32 v13, s12
+; GFX11-NEXT:    v_mov_b32_e32 v15, s9
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB105_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v18
+; GFX11-NEXT:    v_mov_b32_e32 v4, v19
+; GFX11-NEXT:    v_mov_b32_e32 v8, v16
+; GFX11-NEXT:    v_mov_b32_e32 v12, v17
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -24956,307 +24498,152 @@ define inreg <16 x i8> @bitcast_v8bf16_to_v16i8_scalar(<8 x bfloat> inreg %a, i3
 ; GFX9-NEXT:    v_mov_b32_e32 v12, v16
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v8bf16_to_v16i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB109_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s18, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB109_4
-; GFX11-TRUE16-NEXT:  .LBB109_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s1, 16
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s4, 0, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s3, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v2, v8 :: v_dual_add_nc_u32 v7, v7, v3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s2, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v4, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v7, v9 :: v_dual_and_b32 v1, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v8, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v11, v8
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v12, v10
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v9, v13
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v14, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v11, v15, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v8, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v6, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v7, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v3, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v10
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[9:10]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-TRUE16-NEXT:    s_branch .LBB109_5
-; GFX11-TRUE16-NEXT:  .LBB109_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr16_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr18_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr17_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB109_2
-; GFX11-TRUE16-NEXT:  .LBB109_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s4
-; GFX11-TRUE16-NEXT:  .LBB109_5: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v17.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v16.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v8bf16_to_v16i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s16, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB109_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s18, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB109_4
-; GFX11-FAKE16-NEXT:  .LBB109_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s1, 16
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s4, 0, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s3, 16
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v2, v8 :: v_dual_add_nc_u32 v7, v7, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v7, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s2, 16
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v10, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v12, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v11, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v5
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v10, v14, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v7, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v6, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v4, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v3, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
-; GFX11-FAKE16-NEXT:    s_branch .LBB109_5
-; GFX11-FAKE16-NEXT:  .LBB109_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr17
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr16
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr18
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    s_branch .LBB109_2
-; GFX11-FAKE16-NEXT:  .LBB109_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v17, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s3 :: v_dual_mov_b32 v9, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s18 :: v_dual_mov_b32 v15, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v13, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s17 :: v_dual_mov_b32 v1, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s16 :: v_dual_mov_b32 v7, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s13 :: v_dual_mov_b32 v5, s10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB109_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v17
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v16
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v8bf16_to_v16i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s16, 0
+; GFX11-NEXT:    s_mov_b32 s8, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB109_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s11, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s18, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s16, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s9, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s17, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s13, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
+; GFX11-NEXT:    s_cbranch_vccnz .LBB109_4
+; GFX11-NEXT:  .LBB109_2: ; %cmp.true
+; GFX11-NEXT:    s_lshl_b32 s4, s1, 16
+; GFX11-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
+; GFX11-NEXT:    s_pack_lh_b32_b16 s4, 0, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v1
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v0
+; GFX11-NEXT:    v_bfe_u32 v7, v3, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v3
+; GFX11-NEXT:    s_lshl_b32 s0, s3, 16
+; GFX11-NEXT:    s_pack_lh_b32_b16 s1, 0, s3
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v2, v8 :: v_dual_add_nc_u32 v7, v7, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
+; GFX11-NEXT:    v_bfe_u32 v2, v5, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v4, v6, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s1
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v7, v9, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s2, 16
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
+; GFX11-NEXT:    s_pack_lh_b32_b16 s0, 0, s2
+; GFX11-NEXT:    v_bfe_u32 v10, v7, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
+; GFX11-NEXT:    v_or_b32_e32 v13, 0x400000, v7
+; GFX11-NEXT:    v_bfe_u32 v12, v9, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v10, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v2
+; GFX11-NEXT:    v_bfe_u32 v2, v3, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, v12, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v3
+; GFX11-NEXT:    v_bfe_u32 v8, v11, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v5
+; GFX11-NEXT:    v_cndmask_b32_e32 v9, v10, v14, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v17
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v11
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v16
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v2, v12, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v2, v6, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v1, v4, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v10, v14, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v10
+; GFX11-NEXT:    v_lshl_or_b32 v9, v3, 16, v7
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
+; GFX11-NEXT:    s_branch .LBB109_5
+; GFX11-NEXT:  .LBB109_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr17
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr16
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr18
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    s_branch .LBB109_2
+; GFX11-NEXT:  .LBB109_4:
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v17, s1
+; GFX11-NEXT:    v_dual_mov_b32 v16, s3 :: v_dual_mov_b32 v9, s15
+; GFX11-NEXT:    v_dual_mov_b32 v14, s18 :: v_dual_mov_b32 v15, s11
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v13, s14
+; GFX11-NEXT:    v_dual_mov_b32 v6, s17 :: v_dual_mov_b32 v1, s12
+; GFX11-NEXT:    v_dual_mov_b32 v10, s16 :: v_dual_mov_b32 v7, s9
+; GFX11-NEXT:    v_dual_mov_b32 v2, s13 :: v_dual_mov_b32 v5, s10
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB109_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v4, v17
+; GFX11-NEXT:    v_mov_b32_e32 v12, v16
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll
index 7556d355a3844..2c78e34823742 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll
@@ -160,44 +160,24 @@ define inreg half @bitcast_i16_to_f16_scalar(i16 inreg %a, i32 inreg %b) {
 ; GFX9-NEXT:  .LBB1_4:
 ; GFX9-NEXT:    s_branch .LBB1_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_i16_to_f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB1_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s0
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB1_3
-; GFX11-TRUE16-NEXT:  .LBB1_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s0, 3
-; GFX11-TRUE16-NEXT:  .LBB1_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB1_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr1_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB1_2
-;
-; GFX11-FAKE16-LABEL: bitcast_i16_to_f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB1_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_cbranch_vccz .LBB1_4
-; GFX11-FAKE16-NEXT:  ; %bb.2: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB1_3:
-; GFX11-FAKE16-NEXT:  .LBB1_4: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_i16_to_f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB1_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
+; GFX11-NEXT:    s_cbranch_vccz .LBB1_4
+; GFX11-NEXT:  ; %bb.2: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB1_3:
+; GFX11-NEXT:  .LBB1_4: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -368,18 +348,16 @@ define inreg i16 @bitcast_f16_to_i16_scalar(half inreg %a, i32 inreg %b) {
 ; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s1, 0
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB3_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
+; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB3_4
 ; GFX11-TRUE16-NEXT:  .LBB3_2: ; %cmp.true
 ; GFX11-TRUE16-NEXT:    v_add_f16_e64 v0.l, 0x200, s0
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ; GFX11-TRUE16-NEXT:  .LBB3_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
 ; GFX11-TRUE16-NEXT:    s_branch .LBB3_2
 ; GFX11-TRUE16-NEXT:  .LBB3_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: bitcast_f16_to_i16_scalar:
@@ -560,44 +538,24 @@ define inreg bfloat @bitcast_i16_to_bf16_scalar(i16 inreg %a, i32 inreg %b) {
 ; GFX9-NEXT:  .LBB5_4:
 ; GFX9-NEXT:    s_branch .LBB5_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_i16_to_bf16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB5_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s0
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB5_3
-; GFX11-TRUE16-NEXT:  .LBB5_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s0, 3
-; GFX11-TRUE16-NEXT:  .LBB5_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB5_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr1_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB5_2
-;
-; GFX11-FAKE16-LABEL: bitcast_i16_to_bf16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB5_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_cbranch_vccz .LBB5_4
-; GFX11-FAKE16-NEXT:  ; %bb.2: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB5_3:
-; GFX11-FAKE16-NEXT:  .LBB5_4: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_i16_to_bf16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB5_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
+; GFX11-NEXT:    s_cbranch_vccz .LBB5_4
+; GFX11-NEXT:  ; %bb.2: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB5_3:
+; GFX11-NEXT:  .LBB5_4: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -827,65 +785,34 @@ define inreg i16 @bitcast_bf16_to_i16_scalar(bfloat inreg %a, i32 inreg %b) {
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s16
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_bf16_to_i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB7_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB7_4
-; GFX11-TRUE16-NEXT:  .LBB7_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB7_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB7_2
-; GFX11-TRUE16-NEXT:  .LBB7_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_bf16_to_i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB7_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB7_4
-; GFX11-FAKE16-NEXT:  .LBB7_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB7_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB7_2
-; GFX11-FAKE16-NEXT:  .LBB7_4:
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_bf16_to_i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB7_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
+; GFX11-NEXT:    s_cbranch_vccnz .LBB7_4
+; GFX11-NEXT:  .LBB7_2: ; %cmp.true
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v1, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB7_3:
+; GFX11-NEXT:    s_branch .LBB7_2
+; GFX11-NEXT:  .LBB7_4:
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -1071,18 +998,16 @@ define inreg bfloat @bitcast_f16_to_bf16_scalar(half inreg %a, i32 inreg %b) {
 ; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s1, 0
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB9_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
+; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB9_4
 ; GFX11-TRUE16-NEXT:  .LBB9_2: ; %cmp.true
 ; GFX11-TRUE16-NEXT:    v_add_f16_e64 v0.l, 0x200, s0
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ; GFX11-TRUE16-NEXT:  .LBB9_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
 ; GFX11-TRUE16-NEXT:    s_branch .LBB9_2
 ; GFX11-TRUE16-NEXT:  .LBB9_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: bitcast_f16_to_bf16_scalar:
@@ -1335,65 +1260,34 @@ define inreg half @bitcast_bf16_to_f16_scalar(bfloat inreg %a, i32 inreg %b) {
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s16
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_bf16_to_f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB11_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB11_4
-; GFX11-TRUE16-NEXT:  .LBB11_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB11_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB11_2
-; GFX11-TRUE16-NEXT:  .LBB11_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_bf16_to_f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB11_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB11_4
-; GFX11-FAKE16-NEXT:  .LBB11_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB11_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB11_2
-; GFX11-FAKE16-NEXT:  .LBB11_4:
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_bf16_to_f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB11_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
+; GFX11-NEXT:    s_cbranch_vccnz .LBB11_4
+; GFX11-NEXT:  .LBB11_2: ; %cmp.true
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v1, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB11_3:
+; GFX11-NEXT:    s_branch .LBB11_2
+; GFX11-NEXT:  .LBB11_4:
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -1410,5 +1304,3 @@ end:
   %phi = phi half [ %a2, %cmp.true ], [ %a3, %cmp.false ]
   ret half %phi
 }
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; GFX11: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll
index 8dc00701dcfd6..a19567bbe24f6 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll
@@ -5618,244 +5618,117 @@ define inreg <32 x i8> @bitcast_v8i32_to_v32i8_scalar(<8 x i32> inreg %a, i32 in
 ; GFX9-NEXT:    ; implicit-def: $sgpr14
 ; GFX9-NEXT:    s_branch .LBB25_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v8i32_to_v32i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB25_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB25_3
-; GFX11-TRUE16-NEXT:  .LBB25_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB25_3: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s12
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB25_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr21_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr20_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB25_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v8i32_to_v32i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB25_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB25_3
-; GFX11-FAKE16-NEXT:  .LBB25_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB25_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s45
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s44 :: v_dual_mov_b32 v3, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s41
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s29 :: v_dual_mov_b32 v11, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s3 :: v_dual_mov_b32 v13, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s27 :: v_dual_mov_b32 v15, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v19, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s17 :: v_dual_mov_b32 v21, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v25, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s15 :: v_dual_mov_b32 v27, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s19 :: v_dual_mov_b32 v29, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s13 :: v_dual_mov_b32 v31, s12
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB25_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr21
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr20
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    s_branch .LBB25_2
+; GFX11-LABEL: bitcast_v8i32_to_v32i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX11-NEXT:    s_mov_b32 s46, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB25_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s12, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s13, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s20, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s21, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s22, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s23, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s24, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s26, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s27, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s28, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s29, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s41, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-NEXT:    s_cbranch_vccnz .LBB25_3
+; GFX11-NEXT:  .LBB25_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s12, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s13, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s20, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s21, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s22, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s23, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s24, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s26, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s27, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s28, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s29, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s41, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 8
+; GFX11-NEXT:  .LBB25_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s45
+; GFX11-NEXT:    v_dual_mov_b32 v2, s44 :: v_dual_mov_b32 v3, s10
+; GFX11-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s43
+; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s41
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s40
+; GFX11-NEXT:    v_dual_mov_b32 v10, s29 :: v_dual_mov_b32 v11, s8
+; GFX11-NEXT:    v_dual_mov_b32 v12, s3 :: v_dual_mov_b32 v13, s28
+; GFX11-NEXT:    v_dual_mov_b32 v14, s27 :: v_dual_mov_b32 v15, s26
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s25
+; GFX11-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v19, s6
+; GFX11-NEXT:    v_dual_mov_b32 v20, s17 :: v_dual_mov_b32 v21, s23
+; GFX11-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s21
+; GFX11-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v25, s20
+; GFX11-NEXT:    v_dual_mov_b32 v26, s15 :: v_dual_mov_b32 v27, s4
+; GFX11-NEXT:    v_dual_mov_b32 v28, s19 :: v_dual_mov_b32 v29, s14
+; GFX11-NEXT:    v_dual_mov_b32 v30, s13 :: v_dual_mov_b32 v31, s12
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB25_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr21
+; GFX11-NEXT:    ; implicit-def: $sgpr20
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    s_branch .LBB25_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -7389,401 +7262,200 @@ define inreg <8 x i32> @bitcast_v32i8_to_v8i32_scalar(<32 x i8> inreg %a, i32 in
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
 ; GFX9-NEXT:    s_branch .LBB27_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32i8_to_v8i32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v7 :: v_dual_mov_b32 v15, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v17, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v3 :: v_dual_mov_b32 v19, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v21, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB27_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s9, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v21
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v19
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v18
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v14, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v24, v25
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v14, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB27_3
-; GFX11-TRUE16-NEXT:  .LBB27_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v21
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v20
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v19
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v9, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v10, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v11, v8
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB27_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB27_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-TRUE16-NEXT:    s_branch .LBB27_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v32i8_to_v8i32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v17, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v2 :: v_dual_mov_b32 v15, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB27_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s17, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v17
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s6, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v14
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s9, 0xffff
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, s6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v20
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v22, v23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB27_3
-; GFX11-FAKE16-NEXT:  .LBB27_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s19, 8
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s21, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v16
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v17
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v18
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v21, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v14, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v19, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v9, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v11, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v15
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v13, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_and_b32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v20, v0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v9
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:  .LBB27_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB27_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-FAKE16-NEXT:    s_branch .LBB27_2
+; GFX11-LABEL: bitcast_v32i8_to_v8i32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v17, v4
+; GFX11-NEXT:    v_dual_mov_b32 v16, v2 :: v_dual_mov_b32 v15, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB27_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_lshl_b32 s7, s17, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_and_b32 s6, s16, 0xff
+; GFX11-NEXT:    s_and_b32 s8, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s19, 8
+; GFX11-NEXT:    s_or_b32 s6, s6, s7
+; GFX11-NEXT:    s_or_b32 s7, s8, s9
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v16
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v17
+; GFX11-NEXT:    s_and_b32 s6, s6, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_and_b32 s8, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s21, 8
+; GFX11-NEXT:    s_or_b32 s6, s6, s7
+; GFX11-NEXT:    s_or_b32 s7, s8, s9
+; GFX11-NEXT:    s_and_b32 s8, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s23, 8
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v21
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v10
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v14
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v15
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v8
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v12
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    s_and_b32 s8, s9, 0xffff
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v19
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v9
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v13
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-NEXT:    v_mov_b32_e32 v1, s6
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v20
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_and_b32 s10, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s9, 16
+; GFX11-NEXT:    s_lshl_b32 s11, s29, 8
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-NEXT:    v_or_b32_e32 v6, v3, v7
+; GFX11-NEXT:    s_and_b32 s10, s10, 0xffff
+; GFX11-NEXT:    v_or_b32_e32 v7, v22, v23
+; GFX11-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
+; GFX11-NEXT:    v_or_b32_e32 v4, s10, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, s5
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB27_3
+; GFX11-NEXT:  .LBB27_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s17, 8
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s19, 8
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s21, 8
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-NEXT:    s_and_b32 s3, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s23, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v16
+; GFX11-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v17
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v18
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v10
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    v_or_b32_e32 v1, v21, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v12
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_or_b32_e32 v2, v14, v2
+; GFX11-NEXT:    s_and_b32 s4, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v3, v19, v3
+; GFX11-NEXT:    v_or_b32_e32 v4, v9, v4
+; GFX11-NEXT:    v_or_b32_e32 v5, v11, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v15
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_or_b32_e32 v6, v13, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 16
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v5
+; GFX11-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_and_b32 v0, 0xff, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v3, v7
+; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    v_or_b32_e32 v0, v20, v0
+; GFX11-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s29, 8
+; GFX11-NEXT:    v_or_b32_e32 v7, v8, v9
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    v_mov_b32_e32 v2, s2
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:  .LBB27_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB27_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+; GFX11-NEXT:    s_branch .LBB27_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -12967,268 +12639,130 @@ define inreg <32 x i8> @bitcast_v8f32_to_v32i8_scalar(<8 x float> inreg %a, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v25, v32
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v8f32_to_v32i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, s1, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, s3, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, s17, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v28, s19, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v27, s18, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, s16, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, s2, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, s0, 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[32:33], 24, v[27:28]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[24:25], 24, v[19:20]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_5
-; GFX11-TRUE16-NEXT:  .LBB49_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr21_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr20_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s4
-; GFX11-TRUE16-NEXT:  .LBB49_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v16.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v19.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v24.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v27.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v8f32_to_v32i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s12, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v39, s1, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v37, s3, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v35, s17, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v33, s19, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, s18, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, s16, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v36, s2, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v38, s0, 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_5
-; GFX11-FAKE16-NEXT:  .LBB49_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr21
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr20
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s0 :: v_dual_mov_b32 v39, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s2 :: v_dual_mov_b32 v37, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s16 :: v_dual_mov_b32 v35, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s18 :: v_dual_mov_b32 v33, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v2, s44
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s45 :: v_dual_mov_b32 v6, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s42 :: v_dual_mov_b32 v10, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s41 :: v_dual_mov_b32 v14, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v18, s24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v22, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v26, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v21, s25 :: v_dual_mov_b32 v30, s14
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v23, s22
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v25, s21
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v29, s20
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v31, s13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v27, s10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v19, s8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB49_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v38
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v39
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v36
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v37
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v16, v34
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v20, v35
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v24, v32
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v28, v33
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v8f32_to_v32i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX11-NEXT:    s_mov_b32 s12, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s13, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s20, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s21, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s22, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s23, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s24, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s26, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s27, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s28, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s29, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_4
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f32_e64 v39, s1, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v37, s3, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v35, s17, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v33, s19, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v32, s18, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v34, s16, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v36, s2, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v38, s0, 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 24, v33
+; GFX11-NEXT:    v_lshrrev_b64 v[27:28], 24, v[32:33]
+; GFX11-NEXT:    v_lshrrev_b64 v[19:20], 24, v[34:35]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[36:37]
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[38:39]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 8, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 8, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 24, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 8, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 8, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v38
+; GFX11-NEXT:    s_branch .LBB49_5
+; GFX11-NEXT:  .LBB49_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr21
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr20
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    s_branch .LBB49_2
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    v_dual_mov_b32 v38, s0 :: v_dual_mov_b32 v39, s1
+; GFX11-NEXT:    v_dual_mov_b32 v36, s2 :: v_dual_mov_b32 v37, s3
+; GFX11-NEXT:    v_dual_mov_b32 v34, s16 :: v_dual_mov_b32 v35, s17
+; GFX11-NEXT:    v_dual_mov_b32 v32, s18 :: v_dual_mov_b32 v33, s19
+; GFX11-NEXT:    v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v2, s44
+; GFX11-NEXT:    v_dual_mov_b32 v5, s45 :: v_dual_mov_b32 v6, s43
+; GFX11-NEXT:    v_dual_mov_b32 v7, s42 :: v_dual_mov_b32 v10, s29
+; GFX11-NEXT:    v_dual_mov_b32 v9, s41 :: v_dual_mov_b32 v14, s28
+; GFX11-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v18, s24
+; GFX11-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v22, s23
+; GFX11-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v26, s15
+; GFX11-NEXT:    v_dual_mov_b32 v21, s25 :: v_dual_mov_b32 v30, s14
+; GFX11-NEXT:    v_mov_b32_e32 v23, s22
+; GFX11-NEXT:    v_mov_b32_e32 v25, s21
+; GFX11-NEXT:    v_mov_b32_e32 v29, s20
+; GFX11-NEXT:    v_mov_b32_e32 v31, s13
+; GFX11-NEXT:    v_mov_b32_e32 v27, s10
+; GFX11-NEXT:    v_mov_b32_e32 v19, s8
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB49_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v38
+; GFX11-NEXT:    v_mov_b32_e32 v4, v39
+; GFX11-NEXT:    v_mov_b32_e32 v8, v36
+; GFX11-NEXT:    v_mov_b32_e32 v12, v37
+; GFX11-NEXT:    v_mov_b32_e32 v16, v34
+; GFX11-NEXT:    v_mov_b32_e32 v20, v35
+; GFX11-NEXT:    v_mov_b32_e32 v24, v32
+; GFX11-NEXT:    v_mov_b32_e32 v28, v33
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -14762,401 +14296,200 @@ define inreg <8 x float> @bitcast_v32i8_to_v8f32_scalar(<32 x i8> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
 ; GFX9-NEXT:    s_branch .LBB51_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32i8_to_v8f32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v7 :: v_dual_mov_b32 v15, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v17, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v3 :: v_dual_mov_b32 v19, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v21, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB51_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s9, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v21
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v19
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v18
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v14, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v24, v25
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v14, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB51_3
-; GFX11-TRUE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v21
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v20
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v19
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v9, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v10, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v11, v8
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB51_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB51_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-TRUE16-NEXT:    s_branch .LBB51_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v32i8_to_v8f32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v17, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v2 :: v_dual_mov_b32 v15, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB51_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s17, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v17
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s6, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v14
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s9, 0xffff
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, s6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v20
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v22, v23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB51_3
-; GFX11-FAKE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s19, 8
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s21, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v16
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v17
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v18
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v21, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v14, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v19, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v9, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v11, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v15
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v13, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_and_b32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v20, v0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v9
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:  .LBB51_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB51_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-FAKE16-NEXT:    s_branch .LBB51_2
+; GFX11-LABEL: bitcast_v32i8_to_v8f32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v17, v4
+; GFX11-NEXT:    v_dual_mov_b32 v16, v2 :: v_dual_mov_b32 v15, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB51_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_lshl_b32 s7, s17, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_and_b32 s6, s16, 0xff
+; GFX11-NEXT:    s_and_b32 s8, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s19, 8
+; GFX11-NEXT:    s_or_b32 s6, s6, s7
+; GFX11-NEXT:    s_or_b32 s7, s8, s9
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v16
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v17
+; GFX11-NEXT:    s_and_b32 s6, s6, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_and_b32 s8, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s21, 8
+; GFX11-NEXT:    s_or_b32 s6, s6, s7
+; GFX11-NEXT:    s_or_b32 s7, s8, s9
+; GFX11-NEXT:    s_and_b32 s8, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s23, 8
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v21
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v10
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v14
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v15
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v8
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v12
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    s_and_b32 s8, s9, 0xffff
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v19
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v9
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v13
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-NEXT:    v_mov_b32_e32 v1, s6
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v20
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_and_b32 s10, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s9, 16
+; GFX11-NEXT:    s_lshl_b32 s11, s29, 8
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-NEXT:    v_or_b32_e32 v6, v3, v7
+; GFX11-NEXT:    s_and_b32 s10, s10, 0xffff
+; GFX11-NEXT:    v_or_b32_e32 v7, v22, v23
+; GFX11-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
+; GFX11-NEXT:    v_or_b32_e32 v4, s10, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, s5
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB51_3
+; GFX11-NEXT:  .LBB51_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s17, 8
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s19, 8
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s21, 8
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-NEXT:    s_and_b32 s3, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s23, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v16
+; GFX11-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v17
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v18
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v10
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    v_or_b32_e32 v1, v21, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v12
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_or_b32_e32 v2, v14, v2
+; GFX11-NEXT:    s_and_b32 s4, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v3, v19, v3
+; GFX11-NEXT:    v_or_b32_e32 v4, v9, v4
+; GFX11-NEXT:    v_or_b32_e32 v5, v11, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v15
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_or_b32_e32 v6, v13, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 16
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v5
+; GFX11-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_and_b32 v0, 0xff, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v3, v7
+; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    v_or_b32_e32 v0, v20, v0
+; GFX11-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s29, 8
+; GFX11-NEXT:    v_or_b32_e32 v7, v8, v9
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    v_mov_b32_e32 v2, s2
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:  .LBB51_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB51_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+; GFX11-NEXT:    s_branch .LBB51_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -19871,244 +19204,117 @@ define inreg <32 x i8> @bitcast_v4i64_to_v32i8_scalar(<4 x i64> inreg %a, i32 in
 ; GFX9-NEXT:    ; implicit-def: $sgpr14
 ; GFX9-NEXT:    s_branch .LBB69_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4i64_to_v32i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB69_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB69_3
-; GFX11-TRUE16-NEXT:  .LBB69_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB69_3: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s12
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB69_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr21_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr20_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB69_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v4i64_to_v32i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB69_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB69_3
-; GFX11-FAKE16-NEXT:  .LBB69_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB69_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s45
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s44 :: v_dual_mov_b32 v3, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s41
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s29 :: v_dual_mov_b32 v11, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s3 :: v_dual_mov_b32 v13, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s27 :: v_dual_mov_b32 v15, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v19, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s17 :: v_dual_mov_b32 v21, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v25, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s15 :: v_dual_mov_b32 v27, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s19 :: v_dual_mov_b32 v29, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s13 :: v_dual_mov_b32 v31, s12
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB69_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr21
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr20
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    s_branch .LBB69_2
+; GFX11-LABEL: bitcast_v4i64_to_v32i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX11-NEXT:    s_mov_b32 s46, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB69_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s12, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s13, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s20, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s21, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s22, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s23, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s24, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s26, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s27, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s28, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s29, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s41, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-NEXT:    s_cbranch_vccnz .LBB69_3
+; GFX11-NEXT:  .LBB69_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s12, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s13, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s20, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s21, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s22, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s23, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s24, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s26, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s27, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s28, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s29, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s41, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 8
+; GFX11-NEXT:  .LBB69_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s45
+; GFX11-NEXT:    v_dual_mov_b32 v2, s44 :: v_dual_mov_b32 v3, s10
+; GFX11-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s43
+; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s41
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s40
+; GFX11-NEXT:    v_dual_mov_b32 v10, s29 :: v_dual_mov_b32 v11, s8
+; GFX11-NEXT:    v_dual_mov_b32 v12, s3 :: v_dual_mov_b32 v13, s28
+; GFX11-NEXT:    v_dual_mov_b32 v14, s27 :: v_dual_mov_b32 v15, s26
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s25
+; GFX11-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v19, s6
+; GFX11-NEXT:    v_dual_mov_b32 v20, s17 :: v_dual_mov_b32 v21, s23
+; GFX11-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s21
+; GFX11-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v25, s20
+; GFX11-NEXT:    v_dual_mov_b32 v26, s15 :: v_dual_mov_b32 v27, s4
+; GFX11-NEXT:    v_dual_mov_b32 v28, s19 :: v_dual_mov_b32 v29, s14
+; GFX11-NEXT:    v_dual_mov_b32 v30, s13 :: v_dual_mov_b32 v31, s12
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB69_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr21
+; GFX11-NEXT:    ; implicit-def: $sgpr20
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    s_branch .LBB69_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -21642,401 +20848,200 @@ define inreg <4 x i64> @bitcast_v32i8_to_v4i64_scalar(<32 x i8> inreg %a, i32 in
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
 ; GFX9-NEXT:    s_branch .LBB71_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32i8_to_v4i64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v7 :: v_dual_mov_b32 v15, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v17, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v3 :: v_dual_mov_b32 v19, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v21, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB71_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s9, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v21
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v19
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v18
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v14, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v24, v25
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v14, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB71_3
-; GFX11-TRUE16-NEXT:  .LBB71_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v21
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v20
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v19
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v9, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v10, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v11, v8
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB71_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB71_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-TRUE16-NEXT:    s_branch .LBB71_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v32i8_to_v4i64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v17, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v2 :: v_dual_mov_b32 v15, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB71_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s17, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v17
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s6, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v14
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s9, 0xffff
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, s6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v20
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v22, v23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB71_3
-; GFX11-FAKE16-NEXT:  .LBB71_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s19, 8
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s21, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v16
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v17
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v18
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v21, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v14, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v19, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v9, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v11, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v15
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v13, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_and_b32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v20, v0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v9
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:  .LBB71_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB71_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-FAKE16-NEXT:    s_branch .LBB71_2
+; GFX11-LABEL: bitcast_v32i8_to_v4i64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v17, v4
+; GFX11-NEXT:    v_dual_mov_b32 v16, v2 :: v_dual_mov_b32 v15, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB71_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_lshl_b32 s7, s17, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_and_b32 s6, s16, 0xff
+; GFX11-NEXT:    s_and_b32 s8, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s19, 8
+; GFX11-NEXT:    s_or_b32 s6, s6, s7
+; GFX11-NEXT:    s_or_b32 s7, s8, s9
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v16
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v17
+; GFX11-NEXT:    s_and_b32 s6, s6, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_and_b32 s8, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s21, 8
+; GFX11-NEXT:    s_or_b32 s6, s6, s7
+; GFX11-NEXT:    s_or_b32 s7, s8, s9
+; GFX11-NEXT:    s_and_b32 s8, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s23, 8
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v21
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v10
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v14
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v15
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v8
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v12
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    s_and_b32 s8, s9, 0xffff
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v19
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v9
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v13
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-NEXT:    v_mov_b32_e32 v1, s6
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v20
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_and_b32 s10, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s9, 16
+; GFX11-NEXT:    s_lshl_b32 s11, s29, 8
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-NEXT:    v_or_b32_e32 v6, v3, v7
+; GFX11-NEXT:    s_and_b32 s10, s10, 0xffff
+; GFX11-NEXT:    v_or_b32_e32 v7, v22, v23
+; GFX11-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
+; GFX11-NEXT:    v_or_b32_e32 v4, s10, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, s5
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB71_3
+; GFX11-NEXT:  .LBB71_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s17, 8
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s19, 8
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s21, 8
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-NEXT:    s_and_b32 s3, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s23, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v16
+; GFX11-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v17
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v18
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v10
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    v_or_b32_e32 v1, v21, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v12
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_or_b32_e32 v2, v14, v2
+; GFX11-NEXT:    s_and_b32 s4, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v3, v19, v3
+; GFX11-NEXT:    v_or_b32_e32 v4, v9, v4
+; GFX11-NEXT:    v_or_b32_e32 v5, v11, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v15
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_or_b32_e32 v6, v13, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 16
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v5
+; GFX11-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_and_b32 v0, 0xff, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v3, v7
+; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    v_or_b32_e32 v0, v20, v0
+; GFX11-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s29, 8
+; GFX11-NEXT:    v_or_b32_e32 v7, v8, v9
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    v_mov_b32_e32 v2, s2
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:  .LBB71_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB71_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+; GFX11-NEXT:    s_branch .LBB71_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -26230,263 +25235,128 @@ define inreg <32 x i8> @bitcast_v4f64_to_v32i8_scalar(<4 x double> inreg %a, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v25, v32
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4f64_to_v32i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB85_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB85_4
-; GFX11-TRUE16-NEXT:  .LBB85_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[27:28], s[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], s[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[11:12], s[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[3:4], s[0:1], 1.0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[32:33], 24, v[27:28]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[24:25], 24, v[19:20]
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB85_5
-; GFX11-TRUE16-NEXT:  .LBB85_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr21_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr20_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB85_2
-; GFX11-TRUE16-NEXT:  .LBB85_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s28
-; GFX11-TRUE16-NEXT:  .LBB85_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v16.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v19.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v24.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v27.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4f64_to_v32i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s12, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB85_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB85_4
-; GFX11-FAKE16-NEXT:  .LBB85_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[32:33], s[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[34:35], s[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[36:37], s[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[38:39], s[0:1], 1.0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-FAKE16-NEXT:    s_branch .LBB85_5
-; GFX11-FAKE16-NEXT:  .LBB85_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr20
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr21
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    s_branch .LBB85_2
-; GFX11-FAKE16-NEXT:  .LBB85_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s0 :: v_dual_mov_b32 v33, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s2 :: v_dual_mov_b32 v35, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s16 :: v_dual_mov_b32 v37, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s18 :: v_dual_mov_b32 v39, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s46 :: v_dual_mov_b32 v1, s45
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s44 :: v_dual_mov_b32 v9, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s42 :: v_dual_mov_b32 v17, s41
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s40 :: v_dual_mov_b32 v25, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_mov_b32 v30, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s6 :: v_dual_mov_b32 v22, s24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s8 :: v_dual_mov_b32 v14, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s10 :: v_dual_mov_b32 v6, s14
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v31, s28
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v29, s26
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v23, s25
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v21, s23
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v15, s22
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v13, s20
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v7, s15
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v5, s13
-; GFX11-FAKE16-NEXT:  .LBB85_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v38
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v39
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v36
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v37
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v16, v34
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v20, v35
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v24, v32
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v28, v33
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4f64_to_v32i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX11-NEXT:    s_mov_b32 s12, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB85_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s28, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s27, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s26, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s40, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s29, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s25, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s24, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s23, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s22, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s21, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s20, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s15, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s14, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
+; GFX11-NEXT:    s_cbranch_vccnz .LBB85_4
+; GFX11-NEXT:  .LBB85_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[32:33], s[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[34:35], s[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[36:37], s[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[38:39], s[0:1], 1.0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b64 v[27:28], 24, v[32:33]
+; GFX11-NEXT:    v_lshrrev_b64 v[19:20], 24, v[34:35]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[36:37]
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[38:39]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 24, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 8, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 8, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 24, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 8, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 8, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v38
+; GFX11-NEXT:    s_branch .LBB85_5
+; GFX11-NEXT:  .LBB85_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr20
+; GFX11-NEXT:    ; implicit-def: $sgpr21
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    s_branch .LBB85_2
+; GFX11-NEXT:  .LBB85_4:
+; GFX11-NEXT:    v_dual_mov_b32 v38, s0 :: v_dual_mov_b32 v33, s19
+; GFX11-NEXT:    v_dual_mov_b32 v36, s2 :: v_dual_mov_b32 v35, s17
+; GFX11-NEXT:    v_dual_mov_b32 v34, s16 :: v_dual_mov_b32 v37, s3
+; GFX11-NEXT:    v_dual_mov_b32 v32, s18 :: v_dual_mov_b32 v39, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s46 :: v_dual_mov_b32 v1, s45
+; GFX11-NEXT:    v_dual_mov_b32 v10, s44 :: v_dual_mov_b32 v9, s43
+; GFX11-NEXT:    v_dual_mov_b32 v18, s42 :: v_dual_mov_b32 v17, s41
+; GFX11-NEXT:    v_dual_mov_b32 v26, s40 :: v_dual_mov_b32 v25, s29
+; GFX11-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_mov_b32 v30, s27
+; GFX11-NEXT:    v_dual_mov_b32 v11, s6 :: v_dual_mov_b32 v22, s24
+; GFX11-NEXT:    v_dual_mov_b32 v19, s8 :: v_dual_mov_b32 v14, s21
+; GFX11-NEXT:    v_dual_mov_b32 v27, s10 :: v_dual_mov_b32 v6, s14
+; GFX11-NEXT:    v_mov_b32_e32 v31, s28
+; GFX11-NEXT:    v_mov_b32_e32 v29, s26
+; GFX11-NEXT:    v_mov_b32_e32 v23, s25
+; GFX11-NEXT:    v_mov_b32_e32 v21, s23
+; GFX11-NEXT:    v_mov_b32_e32 v15, s22
+; GFX11-NEXT:    v_mov_b32_e32 v13, s20
+; GFX11-NEXT:    v_mov_b32_e32 v7, s15
+; GFX11-NEXT:    v_mov_b32_e32 v5, s13
+; GFX11-NEXT:  .LBB85_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v38
+; GFX11-NEXT:    v_mov_b32_e32 v4, v39
+; GFX11-NEXT:    v_mov_b32_e32 v8, v36
+; GFX11-NEXT:    v_mov_b32_e32 v12, v37
+; GFX11-NEXT:    v_mov_b32_e32 v16, v34
+; GFX11-NEXT:    v_mov_b32_e32 v20, v35
+; GFX11-NEXT:    v_mov_b32_e32 v24, v32
+; GFX11-NEXT:    v_mov_b32_e32 v28, v33
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -28020,401 +26890,200 @@ define inreg <4 x double> @bitcast_v32i8_to_v4f64_scalar(<32 x i8> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
 ; GFX9-NEXT:    s_branch .LBB87_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32i8_to_v4f64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v7 :: v_dual_mov_b32 v15, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v17, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v3 :: v_dual_mov_b32 v19, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v2 :: v_dual_mov_b32 v21, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB87_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s17, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s9, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v21
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v19
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v18
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v14, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v24, v25
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v14, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB87_3
-; GFX11-TRUE16-NEXT:  .LBB87_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v21
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v20
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v19
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v9, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v10, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v11, v8
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB87_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB87_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-TRUE16-NEXT:    s_branch .LBB87_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v32i8_to_v4f64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v17, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v2 :: v_dual_mov_b32 v15, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB87_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s17, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v17
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s6, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v14
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s9, 0xffff
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, s6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v20
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v22, v23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB87_3
-; GFX11-FAKE16-NEXT:  .LBB87_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s19, 8
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s21, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v16
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v17
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v18
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v21, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v14, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v19, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v9, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v11, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v15
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v13, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_and_b32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v3, v7
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v20, v0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v9
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:  .LBB87_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB87_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-FAKE16-NEXT:    s_branch .LBB87_2
+; GFX11-LABEL: bitcast_v32i8_to_v4f64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v17, v4
+; GFX11-NEXT:    v_dual_mov_b32 v16, v2 :: v_dual_mov_b32 v15, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB87_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_lshl_b32 s7, s17, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_and_b32 s6, s16, 0xff
+; GFX11-NEXT:    s_and_b32 s8, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s19, 8
+; GFX11-NEXT:    s_or_b32 s6, s6, s7
+; GFX11-NEXT:    s_or_b32 s7, s8, s9
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v16
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v17
+; GFX11-NEXT:    s_and_b32 s6, s6, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_and_b32 s8, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s21, 8
+; GFX11-NEXT:    s_or_b32 s6, s6, s7
+; GFX11-NEXT:    s_or_b32 s7, s8, s9
+; GFX11-NEXT:    s_and_b32 s8, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s23, 8
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v21
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v10
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v14
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v15
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v8
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v12
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    s_and_b32 s8, s9, 0xffff
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v19
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v9
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v13
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-NEXT:    v_mov_b32_e32 v1, s6
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v20
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_and_b32 s10, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s9, 16
+; GFX11-NEXT:    s_lshl_b32 s11, s29, 8
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v6
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-NEXT:    v_or_b32_e32 v6, v3, v7
+; GFX11-NEXT:    s_and_b32 s10, s10, 0xffff
+; GFX11-NEXT:    v_or_b32_e32 v7, v22, v23
+; GFX11-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
+; GFX11-NEXT:    v_or_b32_e32 v4, s10, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, s5
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB87_3
+; GFX11-NEXT:  .LBB87_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s17, 8
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s19, 8
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s21, 8
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-NEXT:    s_and_b32 s3, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s23, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v16
+; GFX11-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v17
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v18
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v10
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    v_or_b32_e32 v1, v21, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v12
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_or_b32_e32 v2, v14, v2
+; GFX11-NEXT:    s_and_b32 s4, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v3, v19, v3
+; GFX11-NEXT:    v_or_b32_e32 v4, v9, v4
+; GFX11-NEXT:    v_or_b32_e32 v5, v11, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v15
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_or_b32_e32 v6, v13, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 16
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v5
+; GFX11-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_and_b32 v0, 0xff, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v3, v7
+; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    v_or_b32_e32 v0, v20, v0
+; GFX11-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s29, 8
+; GFX11-NEXT:    v_or_b32_e32 v7, v8, v9
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    v_mov_b32_e32 v2, s2
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:  .LBB87_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB87_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+; GFX11-NEXT:    s_branch .LBB87_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -32299,268 +30968,130 @@ define inreg <32 x i8> @bitcast_v16i16_to_v32i8_scalar(<16 x i16> inreg %a, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v25, v32
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v16i16_to_v32i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB97_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB97_4
-; GFX11-TRUE16-NEXT:  .LBB97_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, s17, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v28, s19, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v27, s18, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, s16, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[32:33], 24, v[27:28]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[24:25], 24, v[19:20]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB97_5
-; GFX11-TRUE16-NEXT:  .LBB97_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr21_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr20_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB97_2
-; GFX11-TRUE16-NEXT:  .LBB97_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s4
-; GFX11-TRUE16-NEXT:  .LBB97_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v16.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v19.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v24.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v27.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v16i16_to_v32i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s12, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB97_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB97_4
-; GFX11-FAKE16-NEXT:  .LBB97_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v39, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v37, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v35, s17, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v33, s19, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v32, s18, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v34, s16, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v36, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v38, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-FAKE16-NEXT:    s_branch .LBB97_5
-; GFX11-FAKE16-NEXT:  .LBB97_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr21
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr20
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    s_branch .LBB97_2
-; GFX11-FAKE16-NEXT:  .LBB97_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s0 :: v_dual_mov_b32 v39, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s2 :: v_dual_mov_b32 v37, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s16 :: v_dual_mov_b32 v35, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s18 :: v_dual_mov_b32 v33, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v2, s44
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s45 :: v_dual_mov_b32 v6, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s42 :: v_dual_mov_b32 v10, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s41 :: v_dual_mov_b32 v14, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v18, s24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v22, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v26, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v21, s25 :: v_dual_mov_b32 v30, s14
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v23, s22
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v25, s21
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v29, s20
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v31, s13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v27, s10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v19, s8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB97_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v38
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v39
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v36
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v37
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v16, v34
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v20, v35
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v24, v32
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v28, v33
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v16i16_to_v32i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX11-NEXT:    s_mov_b32 s12, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB97_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s13, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s20, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s21, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s22, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s23, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s24, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s26, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s27, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s28, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s29, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
+; GFX11-NEXT:    s_cbranch_vccnz .LBB97_4
+; GFX11-NEXT:  .LBB97_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v39, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v37, s3, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v35, s17, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v33, s19, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v32, s18, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v34, s16, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v36, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v38, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 24, v33
+; GFX11-NEXT:    v_lshrrev_b64 v[27:28], 24, v[32:33]
+; GFX11-NEXT:    v_lshrrev_b64 v[19:20], 24, v[34:35]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[36:37]
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[38:39]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 8, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 8, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 24, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 8, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 8, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v38
+; GFX11-NEXT:    s_branch .LBB97_5
+; GFX11-NEXT:  .LBB97_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr21
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr20
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    s_branch .LBB97_2
+; GFX11-NEXT:  .LBB97_4:
+; GFX11-NEXT:    v_dual_mov_b32 v38, s0 :: v_dual_mov_b32 v39, s1
+; GFX11-NEXT:    v_dual_mov_b32 v36, s2 :: v_dual_mov_b32 v37, s3
+; GFX11-NEXT:    v_dual_mov_b32 v34, s16 :: v_dual_mov_b32 v35, s17
+; GFX11-NEXT:    v_dual_mov_b32 v32, s18 :: v_dual_mov_b32 v33, s19
+; GFX11-NEXT:    v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v2, s44
+; GFX11-NEXT:    v_dual_mov_b32 v5, s45 :: v_dual_mov_b32 v6, s43
+; GFX11-NEXT:    v_dual_mov_b32 v7, s42 :: v_dual_mov_b32 v10, s29
+; GFX11-NEXT:    v_dual_mov_b32 v9, s41 :: v_dual_mov_b32 v14, s28
+; GFX11-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v18, s24
+; GFX11-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v22, s23
+; GFX11-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v26, s15
+; GFX11-NEXT:    v_dual_mov_b32 v21, s25 :: v_dual_mov_b32 v30, s14
+; GFX11-NEXT:    v_mov_b32_e32 v23, s22
+; GFX11-NEXT:    v_mov_b32_e32 v25, s21
+; GFX11-NEXT:    v_mov_b32_e32 v29, s20
+; GFX11-NEXT:    v_mov_b32_e32 v31, s13
+; GFX11-NEXT:    v_mov_b32_e32 v27, s10
+; GFX11-NEXT:    v_mov_b32_e32 v19, s8
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB97_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v38
+; GFX11-NEXT:    v_mov_b32_e32 v4, v39
+; GFX11-NEXT:    v_mov_b32_e32 v8, v36
+; GFX11-NEXT:    v_mov_b32_e32 v12, v37
+; GFX11-NEXT:    v_mov_b32_e32 v16, v34
+; GFX11-NEXT:    v_mov_b32_e32 v20, v35
+; GFX11-NEXT:    v_mov_b32_e32 v24, v32
+; GFX11-NEXT:    v_mov_b32_e32 v28, v33
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -34081,356 +32612,177 @@ define inreg <16 x i16> @bitcast_v32i8_to_v16i16_scalar(<32 x i8> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
 ; GFX9-NEXT:    s_branch .LBB99_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32i8_to_v16i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v7 :: v_dual_mov_b32 v21, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v19, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v4 :: v_dual_mov_b32 v15, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v2 :: v_dual_mov_b32 v17, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB99_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v19
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v6, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v13
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v1, 0xffff, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v23, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v14, 16, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB99_3
-; GFX11-TRUE16-NEXT:  .LBB99_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s25, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s17, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s19, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s10
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s9, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v10
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v15
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v6, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v2, 16, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB99_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB99_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-TRUE16-NEXT:    s_branch .LBB99_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v32i8_to_v16i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v17, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v4 :: v_dual_mov_b32 v15, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB99_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v16
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v9
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e64 v3, 0xffff, s10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v2, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, s6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v7, 16, v23
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v3, s8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB99_3
-; GFX11-FAKE16-NEXT:  .LBB99_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s17, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s10, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s19, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s6, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s9, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v16
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v17
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v11, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v21, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v20, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v13, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v9, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v19, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v14, v6
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e64 v7, 0xffff, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v5, 16, v8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v3, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:  .LBB99_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB99_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-FAKE16-NEXT:    s_branch .LBB99_2
+; GFX11-LABEL: bitcast_v32i8_to_v16i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v17, v0
+; GFX11-NEXT:    v_dual_mov_b32 v18, v4 :: v_dual_mov_b32 v15, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB99_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v16
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v20
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v17
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v8
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v21
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v12
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v19
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    s_and_b32 s11, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s12, s29, 8
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v9
+; GFX11-NEXT:    s_or_b32 s10, s11, s12
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e64 v3, 0xffff, s10
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v13
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v5, v2, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, s6
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v14
+; GFX11-NEXT:    v_lshl_or_b32 v6, v6, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v7, v7, 16, v23
+; GFX11-NEXT:    v_mov_b32_e32 v2, s7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v4, v0, 16, v3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v3, s8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB99_3
+; GFX11-NEXT:  .LBB99_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_lshl_b32 s5, s29, 8
+; GFX11-NEXT:    s_and_b32 s4, s28, 0xff
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    s_and_b32 s5, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    s_and_b32 s6, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s27, 8
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_or_b32 s6, s7, s6
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s7, s8, s7
+; GFX11-NEXT:    s_and_b32 s8, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s23, 8
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_or_b32 s8, s9, s8
+; GFX11-NEXT:    s_and_b32 s9, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s17, 8
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_or_b32 s9, s10, s9
+; GFX11-NEXT:    s_and_b32 s10, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s19, 8
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s10, s11, s10
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-NEXT:    s_addk_i32 s9, 0x300
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_addk_i32 s10, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v16
+; GFX11-NEXT:    s_addk_i32 s7, 0x300
+; GFX11-NEXT:    s_addk_i32 s8, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v12
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v18
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v17
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v8
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_or_b32_e32 v0, v11, v0
+; GFX11-NEXT:    v_or_b32_e32 v2, v21, v2
+; GFX11-NEXT:    v_or_b32_e32 v4, v20, v4
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v1, v13, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_or_b32_e32 v3, v9, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-NEXT:    v_or_b32_e32 v5, v19, v5
+; GFX11-NEXT:    v_or_b32_e32 v6, v14, v6
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    v_and_b32_e64 v7, 0xffff, s4
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v4, v6, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v5, v5, 16, v8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v6, v3, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:  .LBB99_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB99_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+; GFX11-NEXT:    s_branch .LBB99_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -37549,268 +35901,130 @@ define inreg <32 x i8> @bitcast_v16f16_to_v32i8_scalar(<16 x half> inreg %a, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v25, v32
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v16f16_to_v32i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB105_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB105_4
-; GFX11-TRUE16-NEXT:  .LBB105_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v28, 0x200, s19 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v27, 0x200, s18 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[32:33], 24, v[27:28]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[24:25], 24, v[19:20]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB105_5
-; GFX11-TRUE16-NEXT:  .LBB105_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr21_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr20_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB105_2
-; GFX11-TRUE16-NEXT:  .LBB105_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s4
-; GFX11-TRUE16-NEXT:  .LBB105_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v16.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v19.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v24.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v27.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v16f16_to_v32i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s12, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB105_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB105_4
-; GFX11-FAKE16-NEXT:  .LBB105_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v39, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v37, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v35, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v33, 0x200, s19 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v32, 0x200, s18 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v34, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v36, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v38, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[32:33]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[34:35]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[36:37]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[38:39]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v32
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v34
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v37
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v36
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v39
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v38
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v38
-; GFX11-FAKE16-NEXT:    s_branch .LBB105_5
-; GFX11-FAKE16-NEXT:  .LBB105_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr21
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr20
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    s_branch .LBB105_2
-; GFX11-FAKE16-NEXT:  .LBB105_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s0 :: v_dual_mov_b32 v39, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s2 :: v_dual_mov_b32 v37, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s16 :: v_dual_mov_b32 v35, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s18 :: v_dual_mov_b32 v33, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v2, s44
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s45 :: v_dual_mov_b32 v6, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s42 :: v_dual_mov_b32 v10, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s41 :: v_dual_mov_b32 v14, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v18, s24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v22, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v26, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v21, s25 :: v_dual_mov_b32 v30, s14
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v23, s22
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v25, s21
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v29, s20
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v31, s13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v27, s10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v19, s8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB105_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v38
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v39
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v36
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v37
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v16, v34
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v20, v35
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v24, v32
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v28, v33
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v16f16_to_v32i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX11-NEXT:    s_mov_b32 s12, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB105_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s13, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s20, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s21, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s22, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s23, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s24, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s26, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s27, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s28, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s29, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s44, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
+; GFX11-NEXT:    s_cbranch_vccnz .LBB105_4
+; GFX11-NEXT:  .LBB105_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v39, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v37, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v35, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v33, 0x200, s19 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v32, 0x200, s18 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v34, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v36, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v38, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 24, v33
+; GFX11-NEXT:    v_lshrrev_b64 v[27:28], 24, v[32:33]
+; GFX11-NEXT:    v_lshrrev_b64 v[19:20], 24, v[34:35]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[36:37]
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[38:39]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 8, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 8, v32
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 24, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 8, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 8, v34
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v37
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v36
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v39
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v38
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v38
+; GFX11-NEXT:    s_branch .LBB105_5
+; GFX11-NEXT:  .LBB105_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr21
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr20
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    s_branch .LBB105_2
+; GFX11-NEXT:  .LBB105_4:
+; GFX11-NEXT:    v_dual_mov_b32 v38, s0 :: v_dual_mov_b32 v39, s1
+; GFX11-NEXT:    v_dual_mov_b32 v36, s2 :: v_dual_mov_b32 v37, s3
+; GFX11-NEXT:    v_dual_mov_b32 v34, s16 :: v_dual_mov_b32 v35, s17
+; GFX11-NEXT:    v_dual_mov_b32 v32, s18 :: v_dual_mov_b32 v33, s19
+; GFX11-NEXT:    v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v2, s44
+; GFX11-NEXT:    v_dual_mov_b32 v5, s45 :: v_dual_mov_b32 v6, s43
+; GFX11-NEXT:    v_dual_mov_b32 v7, s42 :: v_dual_mov_b32 v10, s29
+; GFX11-NEXT:    v_dual_mov_b32 v9, s41 :: v_dual_mov_b32 v14, s28
+; GFX11-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v18, s24
+; GFX11-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v22, s23
+; GFX11-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v26, s15
+; GFX11-NEXT:    v_dual_mov_b32 v21, s25 :: v_dual_mov_b32 v30, s14
+; GFX11-NEXT:    v_mov_b32_e32 v23, s22
+; GFX11-NEXT:    v_mov_b32_e32 v25, s21
+; GFX11-NEXT:    v_mov_b32_e32 v29, s20
+; GFX11-NEXT:    v_mov_b32_e32 v31, s13
+; GFX11-NEXT:    v_mov_b32_e32 v27, s10
+; GFX11-NEXT:    v_mov_b32_e32 v19, s8
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB105_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v38
+; GFX11-NEXT:    v_mov_b32_e32 v4, v39
+; GFX11-NEXT:    v_mov_b32_e32 v8, v36
+; GFX11-NEXT:    v_mov_b32_e32 v12, v37
+; GFX11-NEXT:    v_mov_b32_e32 v16, v34
+; GFX11-NEXT:    v_mov_b32_e32 v20, v35
+; GFX11-NEXT:    v_mov_b32_e32 v24, v32
+; GFX11-NEXT:    v_mov_b32_e32 v28, v33
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -39271,356 +37485,177 @@ define inreg <16 x half> @bitcast_v32i8_to_v16f16_scalar(<32 x i8> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
 ; GFX9-NEXT:    s_branch .LBB107_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32i8_to_v16f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v7 :: v_dual_mov_b32 v21, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v19, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v4 :: v_dual_mov_b32 v15, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v2 :: v_dual_mov_b32 v17, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB107_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v19
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v6, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v13
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v1, 0xffff, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v23, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v14, 16, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB107_3
-; GFX11-TRUE16-NEXT:  .LBB107_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s25, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s17, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s19, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s10
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s9, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v10
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v15
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v6, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v2, 16, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB107_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB107_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-TRUE16-NEXT:    s_branch .LBB107_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v32i8_to_v16f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v17, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v4 :: v_dual_mov_b32 v15, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB107_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v16
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v9
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e64 v3, 0xffff, s10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v2, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, s6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v7, 16, v23
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v3, s8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB107_3
-; GFX11-FAKE16-NEXT:  .LBB107_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s17, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s10, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s19, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s6, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s9, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v16
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v17
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v11, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v21, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v20, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v13, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v9, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v19, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v14, v6
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e64 v7, 0xffff, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v5, 16, v8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v3, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:  .LBB107_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB107_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-FAKE16-NEXT:    s_branch .LBB107_2
+; GFX11-LABEL: bitcast_v32i8_to_v16f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v17, v0
+; GFX11-NEXT:    v_dual_mov_b32 v18, v4 :: v_dual_mov_b32 v15, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB107_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v16
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v20
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v17
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v8
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v21
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v12
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v19
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    s_and_b32 s11, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s12, s29, 8
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v9
+; GFX11-NEXT:    s_or_b32 s10, s11, s12
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e64 v3, 0xffff, s10
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v13
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v5, v2, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, s6
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v14
+; GFX11-NEXT:    v_lshl_or_b32 v6, v6, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v7, v7, 16, v23
+; GFX11-NEXT:    v_mov_b32_e32 v2, s7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v4, v0, 16, v3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v3, s8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB107_3
+; GFX11-NEXT:  .LBB107_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_lshl_b32 s5, s29, 8
+; GFX11-NEXT:    s_and_b32 s4, s28, 0xff
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    s_and_b32 s5, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    s_and_b32 s6, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s27, 8
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_or_b32 s6, s7, s6
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s7, s8, s7
+; GFX11-NEXT:    s_and_b32 s8, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s23, 8
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_or_b32 s8, s9, s8
+; GFX11-NEXT:    s_and_b32 s9, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s17, 8
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_or_b32 s9, s10, s9
+; GFX11-NEXT:    s_and_b32 s10, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s19, 8
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s10, s11, s10
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-NEXT:    s_addk_i32 s9, 0x300
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_addk_i32 s10, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v16
+; GFX11-NEXT:    s_addk_i32 s7, 0x300
+; GFX11-NEXT:    s_addk_i32 s8, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v12
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v18
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v17
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v8
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_or_b32_e32 v0, v11, v0
+; GFX11-NEXT:    v_or_b32_e32 v2, v21, v2
+; GFX11-NEXT:    v_or_b32_e32 v4, v20, v4
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v1, v13, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_or_b32_e32 v3, v9, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-NEXT:    v_or_b32_e32 v5, v19, v5
+; GFX11-NEXT:    v_or_b32_e32 v6, v14, v6
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    v_and_b32_e64 v7, 0xffff, s4
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v4, v6, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v5, v5, 16, v8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v6, v3, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:  .LBB107_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB107_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+; GFX11-NEXT:    s_branch .LBB107_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -41529,572 +39564,281 @@ define inreg <32 x i8> @bitcast_v16bf16_to_v32i8_scalar(<16 x bfloat> inreg %a,
 ; GFX9-NEXT:    v_mov_b32_e32 v28, v32
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v16bf16_to_v32i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB109_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB109_4
-; GFX11-TRUE16-NEXT:  .LBB109_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v3, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s3, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s2, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v7, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v4, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v10, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v8, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v6, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v10, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v9, v11, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v12, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v13, v0, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s2, 16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v9, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v5
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s17, 16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v9, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v13
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s16, 16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v8, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v15, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, v16, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v17, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v9, v15, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v18, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v16, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v16
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s19, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v9, v15, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v16, v19
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s18, 16
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v16, 16, 1
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v18, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v9, v17, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, v24, v18
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, v15, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v23, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, 0x400000, v18
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v17, v25, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v19, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v16
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v21, v26, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v15, v24, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v22, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v7
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v15, v17, v19, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v30, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v5, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v9, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v1, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[32:33], 24, v[25:26]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[9:10]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-TRUE16-NEXT:    s_branch .LBB109_5
-; GFX11-TRUE16-NEXT:  .LBB109_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr20_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr21_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB109_2
-; GFX11-TRUE16-NEXT:  .LBB109_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s4
-; GFX11-TRUE16-NEXT:  .LBB109_5: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v35.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v34.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v33.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v16bf16_to_v32i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s12, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB109_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB109_4
-; GFX11-FAKE16-NEXT:  .LBB109_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v3, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s3, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s2, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v4, v10, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v8, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v4, v11 :: v_dual_add_nc_u32 v12, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v10, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v6, 16, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v9, v8
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v12, v0, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s2, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s17, 16
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v4, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v34
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s16, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v7, v8 :: v_dual_add_nc_u32 v7, v12, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v13, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v15, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v13, v15
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v7, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v15
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v17, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v16, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v17
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s19, 16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v7, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v13, v17
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s18, 16
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v13, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v4, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v16, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v18, v13
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v20, v16
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v12, v4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v19, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v16
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v15, v21, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v17, v19
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v18, v23, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v19
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v12, v20, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v22, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v15, v17, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v30, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v5, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v1, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v7, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[17:18]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[25:26]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-FAKE16-NEXT:    s_branch .LBB109_5
-; GFX11-FAKE16-NEXT:  .LBB109_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr20
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr21
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    s_branch .LBB109_2
-; GFX11-FAKE16-NEXT:  .LBB109_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v33, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s19 :: v_dual_mov_b32 v35, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s46 :: v_dual_mov_b32 v25, s41
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v31, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v29, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v17, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s3 :: v_dual_mov_b32 v23, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s44 :: v_dual_mov_b32 v21, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v9, s24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s43 :: v_dual_mov_b32 v15, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s42 :: v_dual_mov_b32 v13, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s29 :: v_dual_mov_b32 v1, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s25 :: v_dual_mov_b32 v7, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s21 :: v_dual_mov_b32 v5, s14
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v27, s10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v19, s8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB109_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v35
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v34
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v20, v33
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v28, v32
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v16bf16_to_v32i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX11-NEXT:    s_mov_b32 s12, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB109_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s27, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s46, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s23, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s45, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s26, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s29, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s28, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s15, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s22, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s25, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s24, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s21, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s20, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s12
+; GFX11-NEXT:    s_cbranch_vccnz .LBB109_4
+; GFX11-NEXT:  .LBB109_2: ; %cmp.true
+; GFX11-NEXT:    s_lshl_b32 s4, s1, 16
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s4
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
+; GFX11-NEXT:    s_and_b32 s4, s0, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_f32_e64 v3, 0x40c00000, s4
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_or_b32_e32 v11, 0x400000, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_bfe_u32 v10, v3, 16, 1
+; GFX11-NEXT:    s_and_b32 s1, s3, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s0, s2, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_bfe_u32 v2, v5, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v4, v10, v3
+; GFX11-NEXT:    v_bfe_u32 v10, v7, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v35
+; GFX11-NEXT:    v_bfe_u32 v9, v8, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v0, 0x400000, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v3, v4, v11 :: v_dual_add_nc_u32 v12, 0x7fff, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v10, v7
+; GFX11-NEXT:    v_lshl_or_b32 v2, v6, 16, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v9, v8
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v12, v0, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v5, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s2, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v4, v9, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_bfe_u32 v1, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v4
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s17, 0xffff0000
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v5
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s17, 16
+; GFX11-NEXT:    v_bfe_u32 v8, v4, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_or_b32_e32 v16, 0x400000, v9
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v34
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v4
+; GFX11-NEXT:    v_bfe_u32 v12, v9, 16, 1
+; GFX11-NEXT:    v_lshl_or_b32 v10, v14, 16, v7
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v5
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s16, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v4
+; GFX11-NEXT:    v_bfe_u32 v13, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    s_lshl_b32 s0, s16, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s19, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v4, v7, v8 :: v_dual_add_nc_u32 v7, v12, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, v13, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_bfe_u32 v13, v15, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v12
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, v13, v15
+; GFX11-NEXT:    v_cndmask_b32_e32 v5, v7, v12, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v13
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v15
+; GFX11-NEXT:    v_bfe_u32 v13, v17, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v4, v16, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-NEXT:    v_or_b32_e32 v15, 0x400000, v17
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s19, 16
+; GFX11-NEXT:    v_cndmask_b32_e32 v5, v7, v12, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, v13, v17
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s18, 16
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
+; GFX11-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
+; GFX11-NEXT:    v_bfe_u32 v18, v13, 16, 1
+; GFX11-NEXT:    s_and_b32 s0, s18, 0xffff0000
+; GFX11-NEXT:    v_bfe_u32 v12, v4, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v20, v16, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v15, v18, v13
+; GFX11-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
+; GFX11-NEXT:    v_or_b32_e32 v21, 0x400000, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v18, v20, v16
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, v12, v4
+; GFX11-NEXT:    v_bfe_u32 v17, v19, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
+; GFX11-NEXT:    v_or_b32_e32 v23, 0x400000, v16
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v15, v21, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
+; GFX11-NEXT:    v_or_b32_e32 v20, 0x400000, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v17, v17, v19
+; GFX11-NEXT:    v_cndmask_b32_e32 v16, v18, v23, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v17
+; GFX11-NEXT:    v_or_b32_e32 v17, 0x400000, v19
+; GFX11-NEXT:    v_cndmask_b32_e32 v4, v12, v20, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v18, v22, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v4
+; GFX11-NEXT:    v_cndmask_b32_e32 v12, v15, v17, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v32
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 24, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 16, v12
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v26, v30, 16, v4
+; GFX11-NEXT:    v_lshl_or_b32 v17, v5, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v9, v1, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v1, v3, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v25, v7, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 24, v26
+; GFX11-NEXT:    v_lshrrev_b64 v[19:20], 24, v[17:18]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[27:28], 24, v[25:26]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 8, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 8, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 8, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
+; GFX11-NEXT:    s_branch .LBB109_5
+; GFX11-NEXT:  .LBB109_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr20
+; GFX11-NEXT:    ; implicit-def: $sgpr21
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    s_branch .LBB109_2
+; GFX11-NEXT:  .LBB109_4:
+; GFX11-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v33, s17
+; GFX11-NEXT:    v_dual_mov_b32 v32, s19 :: v_dual_mov_b32 v35, s1
+; GFX11-NEXT:    v_dual_mov_b32 v30, s46 :: v_dual_mov_b32 v25, s41
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v31, s27
+; GFX11-NEXT:    v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v29, s40
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v17, s28
+; GFX11-NEXT:    v_dual_mov_b32 v34, s3 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT:    v_dual_mov_b32 v14, s44 :: v_dual_mov_b32 v21, s26
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v9, s24
+; GFX11-NEXT:    v_dual_mov_b32 v6, s43 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT:    v_dual_mov_b32 v26, s42 :: v_dual_mov_b32 v13, s22
+; GFX11-NEXT:    v_dual_mov_b32 v18, s29 :: v_dual_mov_b32 v1, s20
+; GFX11-NEXT:    v_dual_mov_b32 v10, s25 :: v_dual_mov_b32 v7, s13
+; GFX11-NEXT:    v_dual_mov_b32 v2, s21 :: v_dual_mov_b32 v5, s14
+; GFX11-NEXT:    v_mov_b32_e32 v27, s10
+; GFX11-NEXT:    v_mov_b32_e32 v19, s8
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB109_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v4, v35
+; GFX11-NEXT:    v_mov_b32_e32 v12, v34
+; GFX11-NEXT:    v_mov_b32_e32 v20, v33
+; GFX11-NEXT:    v_mov_b32_e32 v28, v32
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -43609,356 +41353,177 @@ define inreg <16 x bfloat> @bitcast_v32i8_to_v16bf16_scalar(<32 x i8> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
 ; GFX9-NEXT:    s_branch .LBB111_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32i8_to_v16bf16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v7 :: v_dual_mov_b32 v21, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v6 :: v_dual_mov_b32 v19, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v4 :: v_dual_mov_b32 v15, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v2 :: v_dual_mov_b32 v17, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB111_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v19
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v6, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v13
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v1, 0xffff, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v23, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v14, 16, v23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB111_3
-; GFX11-TRUE16-NEXT:  .LBB111_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s25, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s17, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s19, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s10
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s9, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v10
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v15
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v6, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v2, 16, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB111_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB111_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-TRUE16-NEXT:    s_branch .LBB111_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v32i8_to_v16bf16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v17, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, v4 :: v_dual_mov_b32 v15, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB111_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v16
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v9
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e64 v3, 0xffff, s10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v2, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, s6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v7, 16, v23
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v3, s8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB111_3
-; GFX11-FAKE16-NEXT:  .LBB111_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s17, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s10, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s19, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s6, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s9, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v16
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v17
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v11, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v21, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v20, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v13, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v9, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v19, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v14, v6
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e64 v7, 0xffff, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v5, 16, v8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v3, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:  .LBB111_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB111_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
-; GFX11-FAKE16-NEXT:    s_branch .LBB111_2
+; GFX11-LABEL: bitcast_v32i8_to_v16bf16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v14
+; GFX11-NEXT:    v_dual_mov_b32 v16, v6 :: v_dual_mov_b32 v17, v0
+; GFX11-NEXT:    v_dual_mov_b32 v18, v4 :: v_dual_mov_b32 v15, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB111_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v16
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v20
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v17
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v8
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v21
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v12
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v19
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    s_and_b32 s11, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s12, s29, 8
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v9
+; GFX11-NEXT:    s_or_b32 s10, s11, s12
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e64 v3, 0xffff, s10
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v13
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v5, v2, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, s6
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v14
+; GFX11-NEXT:    v_lshl_or_b32 v6, v6, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v7, v7, 16, v23
+; GFX11-NEXT:    v_mov_b32_e32 v2, s7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v4, v0, 16, v3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v3, s8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB111_3
+; GFX11-NEXT:  .LBB111_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_lshl_b32 s5, s29, 8
+; GFX11-NEXT:    s_and_b32 s4, s28, 0xff
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    s_and_b32 s5, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    s_and_b32 s6, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s27, 8
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_or_b32 s6, s7, s6
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s7, s8, s7
+; GFX11-NEXT:    s_and_b32 s8, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s23, 8
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_or_b32 s8, s9, s8
+; GFX11-NEXT:    s_and_b32 s9, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s17, 8
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_or_b32 s9, s10, s9
+; GFX11-NEXT:    s_and_b32 s10, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s19, 8
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s10, s11, s10
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-NEXT:    s_addk_i32 s9, 0x300
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_addk_i32 s10, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v16
+; GFX11-NEXT:    s_addk_i32 s7, 0x300
+; GFX11-NEXT:    s_addk_i32 s8, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v12
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v18
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v17
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v8
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_or_b32_e32 v0, v11, v0
+; GFX11-NEXT:    v_or_b32_e32 v2, v21, v2
+; GFX11-NEXT:    v_or_b32_e32 v4, v20, v4
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v1, v13, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_or_b32_e32 v3, v9, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-NEXT:    v_or_b32_e32 v5, v19, v5
+; GFX11-NEXT:    v_or_b32_e32 v6, v14, v6
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    v_and_b32_e64 v7, 0xffff, s4
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v4, v6, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v5, v5, 16, v8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v6, v3, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v7, v1, 16, v0
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:  .LBB111_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB111_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+; GFX11-NEXT:    s_branch .LBB111_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
index e5a1260fa4538..e773b546afe1b 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
@@ -4072,471 +4072,220 @@ define inreg <40 x i8> @bitcast_v10i32_to_v40i8_scalar(<10 x i32> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $sgpr26
 ; GFX9-NEXT:    s_branch .LBB13_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v10i32_to_v40i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s63
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-TRUE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB13_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s62
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s61
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s12
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s60
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s59
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s58
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s57
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s47
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s46
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s45
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s10, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s44
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s40
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s25
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s14
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-TRUE16-NEXT:    s_clause 0x2
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b64 v0, v[9:10], off offset:32
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB13_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB13_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v10i32_to_v40i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s63, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s63
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-FAKE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB13_3: ; %end
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s62, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s61, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s12, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s60, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s59, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s58, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s57, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s56, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s10, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s47, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s46, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s45, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s44, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s43, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s8, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s42, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s41, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s40, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s29, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s19, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s27, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s25, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s24, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s23, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s4, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s21, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s22, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s15, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s14, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-FAKE16-NEXT:    s_clause 0x2
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b64 v0, v[9:10], off offset:32
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB13_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    s_branch .LBB13_2
+; GFX11-LABEL: bitcast_v10i32_to_v40i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s22, 0
+; GFX11-NEXT:    s_mov_b32 s63, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB13_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s14, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s15, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s22, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s23, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s24, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s25, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s26, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s27, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s28, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s29, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s41, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s59, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s63
+; GFX11-NEXT:    s_cbranch_vccnz .LBB13_3
+; GFX11-NEXT:  .LBB13_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s14, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s15, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s22, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s23, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s24, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s25, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s26, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s27, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s28, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s29, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s41, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s59, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s0, 8
+; GFX11-NEXT:  .LBB13_3: ; %end
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s62, 8
+; GFX11-NEXT:    s_and_b32 s7, s61, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s12, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s5
+; GFX11-NEXT:    s_or_b32 s5, s7, s9
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s60, 8
+; GFX11-NEXT:    s_and_b32 s9, s59, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s58, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s5
+; GFX11-NEXT:    s_or_b32 s1, s1, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s57, 8
+; GFX11-NEXT:    s_and_b32 s7, s56, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s10, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s5
+; GFX11-NEXT:    s_or_b32 s5, s7, s9
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s47, 8
+; GFX11-NEXT:    s_and_b32 s9, s46, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s45, 8
+; GFX11-NEXT:    s_or_b32 s3, s3, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s10
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s5
+; GFX11-NEXT:    s_or_b32 s3, s3, s7
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    s_and_b32 s0, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s44, 8
+; GFX11-NEXT:    s_and_b32 s2, s43, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s8, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s17, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s42, 8
+; GFX11-NEXT:    s_and_b32 s5, s41, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s40, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s29, 8
+; GFX11-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s6
+; GFX11-NEXT:    s_and_b32 s5, s19, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s27, 8
+; GFX11-NEXT:    s_and_b32 s7, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s25, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s6
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    s_and_b32 s0, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s24, 8
+; GFX11-NEXT:    s_and_b32 s2, s23, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s4, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s21, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s22, 8
+; GFX11-NEXT:    s_and_b32 s4, s15, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s14, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
+; GFX11-NEXT:    s_clause 0x2
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-NEXT:    scratch_store_b64 v0, v[9:10], off offset:32
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB13_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    s_branch .LBB13_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -6465,485 +6214,239 @@ define inreg <10 x i32> @bitcast_v40i8_to_v10i32_scalar(<40 x i8> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
 ; GFX9-NEXT:    s_branch .LBB15_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v40i8_to_v10i32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v9 :: v_dual_mov_b32 v23, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v8 :: v_dual_mov_b32 v25, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v5 :: v_dual_mov_b32 v27, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v4 :: v_dual_mov_b32 v29, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v1 :: v_dual_mov_b32 v31, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB15_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v30
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v27
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v22, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v34, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v36, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v34, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB15_3
-; GFX11-TRUE16-NEXT:  .LBB15_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v31
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v30
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v28
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v26
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v6, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v11, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v13, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v14, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v15, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v12, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB15_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB15_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GFX11-TRUE16-NEXT:    s_branch .LBB15_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v40i8_to_v10i32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, v8 :: v_dual_mov_b32 v26, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, v4 :: v_dual_mov_b32 v24, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, v0 :: v_dual_lshlrev_b32 v32, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v21
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB15_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v23
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v27
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v32
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v30
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v14
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v31
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v13
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v25
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v34, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB15_3
-; GFX11-FAKE16-NEXT:  .LBB15_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v26
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v27
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v29, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v30, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v23
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v31, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v13, v6
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v12
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v32, v0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v25
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 3, v16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 3, v18
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 3, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v28, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v11, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v15, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v17, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v19, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v22, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:  .LBB15_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB15_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GFX11-FAKE16-NEXT:    s_branch .LBB15_2
+; GFX11-LABEL: bitcast_v40i8_to_v10i32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
+; GFX11-NEXT:    v_dual_mov_b32 v27, v8 :: v_dual_mov_b32 v26, v6
+; GFX11-NEXT:    v_dual_mov_b32 v25, v4 :: v_dual_mov_b32 v24, v2
+; GFX11-NEXT:    v_dual_mov_b32 v23, v0 :: v_dual_lshlrev_b32 v32, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v30, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v31, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v21
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB15_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v23
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v26
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v27
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v32
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v10
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v29
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v30
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v14
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_and_b32 s8, s8, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s9, s9, 16
+; GFX11-NEXT:    s_and_b32 s10, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s29, 8
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v31
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v13
+; GFX11-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v25
+; GFX11-NEXT:    s_and_b32 s10, s10, 0xffff
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v12
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v16
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v20
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v2, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s8
+; GFX11-NEXT:    v_or_b32_e32 v4, s10, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v24
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v28
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v11
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v15
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v17
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v22
+; GFX11-NEXT:    v_or_b32_e32 v21, v21, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v33
+; GFX11-NEXT:    v_or_b32_e32 v8, v34, v8
+; GFX11-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-NEXT:    v_mov_b32_e32 v0, s5
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v21
+; GFX11-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB15_3
+; GFX11-NEXT:  .LBB15_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s17, 8
+; GFX11-NEXT:    s_and_b32 s3, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_and_b32 s2, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s21, 8
+; GFX11-NEXT:    s_and_b32 s4, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s23, 8
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_or_b32 s3, s5, s4
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v26
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v27
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v14
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    s_and_b32 s3, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_or_b32_e32 v2, v29, v2
+; GFX11-NEXT:    v_or_b32_e32 v3, v30, v3
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v23
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_or_b32_e32 v5, v31, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_or_b32_e32 v6, v13, v6
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v12
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_lshl_b32 s6, s29, 8
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v2, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    v_or_b32_e32 v0, v32, v0
+; GFX11-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v25
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v16
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v18
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v24
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_or_b32_e32 v1, v28, v1
+; GFX11-NEXT:    v_or_b32_e32 v5, v11, v5
+; GFX11-NEXT:    v_or_b32_e32 v8, v15, v8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_or_b32_e32 v9, v17, v9
+; GFX11-NEXT:    v_or_b32_e32 v10, v19, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_or_b32_e32 v0, v22, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v11
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_mov_b32_e32 v2, s2
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v10
+; GFX11-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:  .LBB15_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB15_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
+; GFX11-NEXT:    s_branch .LBB15_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -11533,489 +11036,237 @@ define inreg <40 x i8> @bitcast_v10f32_to_v40i8_scalar(<10 x float> inreg %a, i3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v10f32_to_v40i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB33_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB33_4
-; GFX11-TRUE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, s3, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, s2, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, s1, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, s17, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, s19, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, s21, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, s20, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, s18, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, s16, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v15, s0, 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[9:10], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[13:14], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[18:19], 24, v[15:16]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 24, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 8, v15
-; GFX11-TRUE16-NEXT:    s_branch .LBB33_5
-; GFX11-TRUE16-NEXT:  .LBB33_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB33_2
-; GFX11-TRUE16-NEXT:  .LBB33_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s63
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s61
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s59
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s57
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s47
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s12
-; GFX11-TRUE16-NEXT:  .LBB33_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v39, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v34, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v32, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v29, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v11, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v12, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v37, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v24, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v19, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v10, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v3, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v4, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v6
-; GFX11-TRUE16-NEXT:    s_clause 0x2
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[15:18], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v10f32_to_v40i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB33_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB33_4
-; GFX11-FAKE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, s17, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, s16, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, s3, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, s2, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, s1, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, s19, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, s21, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, s20, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, s18, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, s0, 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 24, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 24, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 8, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 8, v13
-; GFX11-FAKE16-NEXT:    s_branch .LBB33_5
-; GFX11-FAKE16-NEXT:  .LBB33_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    s_branch .LBB33_2
-; GFX11-FAKE16-NEXT:  .LBB33_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v4, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s20 :: v_dual_mov_b32 v2, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v48, s63 :: v_dual_mov_b32 v39, s61
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s62 :: v_dual_mov_b32 v37, s60
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s59 :: v_dual_mov_b32 v35, s58
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s56 :: v_dual_mov_b32 v33, s57
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s47 :: v_dual_mov_b32 v31, s46
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s45 :: v_dual_mov_b32 v29, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s44 :: v_dual_mov_b32 v27, s42
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v25, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s28 :: v_dual_mov_b32 v23, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s27 :: v_dual_mov_b32 v21, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s25 :: v_dual_mov_b32 v19, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v17, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v15, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s15 :: v_dual_mov_b32 v11, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s6 :: v_dual_mov_b32 v7, s12
-; GFX11-FAKE16-NEXT:  .LBB33_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xff, v34
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v39, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v34, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v32, v31
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v29, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v30
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v9, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v10, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v27
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v24, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v22
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v37, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v19, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v3, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v4, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v17
-; GFX11-FAKE16-NEXT:    s_clause 0x2
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v10f32_to_v40i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s22, 0
+; GFX11-NEXT:    s_mov_b32 s14, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB33_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s15, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s22, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s24, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s23, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s26, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s27, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s29, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s28, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s41, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s47, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s59, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
+; GFX11-NEXT:    s_cbranch_vccnz .LBB33_4
+; GFX11-NEXT:  .LBB33_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f32_e64 v6, s17, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v5, s16, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v10, s3, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v9, s2, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v14, s1, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v4, s19, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v2, s21, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v1, s20, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v3, s18, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v13, s0, 1.0
+; GFX11-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b64 v[16:17], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[17:18], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 24, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 24, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 8, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 8, v13
+; GFX11-NEXT:    s_branch .LBB33_5
+; GFX11-NEXT:  .LBB33_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    s_branch .LBB33_2
+; GFX11-NEXT:  .LBB33_4:
+; GFX11-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT:    v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
+; GFX11-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
+; GFX11-NEXT:    v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v4, s19
+; GFX11-NEXT:    v_dual_mov_b32 v1, s20 :: v_dual_mov_b32 v2, s21
+; GFX11-NEXT:    v_dual_mov_b32 v48, s63 :: v_dual_mov_b32 v39, s61
+; GFX11-NEXT:    v_dual_mov_b32 v38, s62 :: v_dual_mov_b32 v37, s60
+; GFX11-NEXT:    v_dual_mov_b32 v36, s59 :: v_dual_mov_b32 v35, s58
+; GFX11-NEXT:    v_dual_mov_b32 v34, s56 :: v_dual_mov_b32 v33, s57
+; GFX11-NEXT:    v_dual_mov_b32 v32, s47 :: v_dual_mov_b32 v31, s46
+; GFX11-NEXT:    v_dual_mov_b32 v30, s45 :: v_dual_mov_b32 v29, s43
+; GFX11-NEXT:    v_dual_mov_b32 v28, s44 :: v_dual_mov_b32 v27, s42
+; GFX11-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v25, s40
+; GFX11-NEXT:    v_dual_mov_b32 v24, s28 :: v_dual_mov_b32 v23, s29
+; GFX11-NEXT:    v_dual_mov_b32 v22, s27 :: v_dual_mov_b32 v21, s26
+; GFX11-NEXT:    v_dual_mov_b32 v20, s25 :: v_dual_mov_b32 v19, s23
+; GFX11-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v17, s4
+; GFX11-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v15, s8
+; GFX11-NEXT:    v_dual_mov_b32 v8, s15 :: v_dual_mov_b32 v11, s10
+; GFX11-NEXT:    v_dual_mov_b32 v16, s6 :: v_dual_mov_b32 v7, s12
+; GFX11-NEXT:  .LBB33_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
+; GFX11-NEXT:    v_and_b32_e32 v39, 0xff, v39
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xff, v34
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v32
+; GFX11-NEXT:    v_lshlrev_b32_e32 v31, 8, v31
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v48
+; GFX11-NEXT:    v_or_b32_e32 v17, v39, v17
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v35
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
+; GFX11-NEXT:    v_or_b32_e32 v16, v34, v16
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v33
+; GFX11-NEXT:    v_or_b32_e32 v31, v32, v31
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
+; GFX11-NEXT:    v_or_b32_e32 v15, v29, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 16, v31
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v30
+; GFX11-NEXT:    v_lshlrev_b32_e32 v30, 16, v15
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v17
+; GFX11-NEXT:    v_or_b32_e32 v15, v9, v16
+; GFX11-NEXT:    v_or_b32_e32 v16, v10, v29
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v28
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v26
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xff, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v9
+; GFX11-NEXT:    v_or_b32_e32 v9, v10, v17
+; GFX11-NEXT:    v_or_b32_e32 v10, v24, v11
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xff, v22
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v21
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v38
+; GFX11-NEXT:    v_or_b32_e32 v36, v37, v36
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v25
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v23
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v17
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v20
+; GFX11-NEXT:    v_or_b32_e32 v7, v19, v7
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v18
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v8
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v35
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v30
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v9
+; GFX11-NEXT:    v_or_b32_e32 v7, v3, v10
+; GFX11-NEXT:    v_or_b32_e32 v8, v4, v11
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v12
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v17
+; GFX11-NEXT:    s_clause 0x2
+; GFX11-NEXT:    scratch_store_b128 v0, v[13:16], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -13944,485 +13195,239 @@ define inreg <10 x float> @bitcast_v40i8_to_v10f32_scalar(<40 x i8> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
 ; GFX9-NEXT:    s_branch .LBB35_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v40i8_to_v10f32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v9 :: v_dual_mov_b32 v23, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v8 :: v_dual_mov_b32 v25, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v5 :: v_dual_mov_b32 v27, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v4 :: v_dual_mov_b32 v29, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v1 :: v_dual_mov_b32 v31, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB35_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v30
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v27
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v22, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v34, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v36, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v34, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB35_3
-; GFX11-TRUE16-NEXT:  .LBB35_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v31
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v30
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v28
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v26
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v6, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v11, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v13, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v14, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v15, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v12, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB35_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB35_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GFX11-TRUE16-NEXT:    s_branch .LBB35_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v40i8_to_v10f32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, v8 :: v_dual_mov_b32 v26, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, v4 :: v_dual_mov_b32 v24, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, v0 :: v_dual_lshlrev_b32 v32, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v21
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB35_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v23
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v27
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v32
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v30
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v14
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v31
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v13
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v25
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v34, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB35_3
-; GFX11-FAKE16-NEXT:  .LBB35_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v26
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v27
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v29, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v30, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v23
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v31, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v13, v6
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v12
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v32, v0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v25
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 3, v16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 3, v18
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 3, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v28, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v11, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v15, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v17, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v19, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v22, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:  .LBB35_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB35_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
-; GFX11-FAKE16-NEXT:    s_branch .LBB35_2
+; GFX11-LABEL: bitcast_v40i8_to_v10f32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
+; GFX11-NEXT:    v_dual_mov_b32 v27, v8 :: v_dual_mov_b32 v26, v6
+; GFX11-NEXT:    v_dual_mov_b32 v25, v4 :: v_dual_mov_b32 v24, v2
+; GFX11-NEXT:    v_dual_mov_b32 v23, v0 :: v_dual_lshlrev_b32 v32, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v30, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v31, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v21
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB35_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v23
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v26
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v27
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v32
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v10
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v29
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v30
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v14
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_and_b32 s8, s8, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s9, s9, 16
+; GFX11-NEXT:    s_and_b32 s10, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s29, 8
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v31
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v13
+; GFX11-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v25
+; GFX11-NEXT:    s_and_b32 s10, s10, 0xffff
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v12
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v16
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xff, v20
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v2, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s8
+; GFX11-NEXT:    v_or_b32_e32 v4, s10, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v24
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v28
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v11
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v15
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v17
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v22
+; GFX11-NEXT:    v_or_b32_e32 v21, v21, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v33
+; GFX11-NEXT:    v_or_b32_e32 v8, v34, v8
+; GFX11-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-NEXT:    v_mov_b32_e32 v0, s5
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v21
+; GFX11-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB35_3
+; GFX11-NEXT:  .LBB35_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s17, 8
+; GFX11-NEXT:    s_and_b32 s3, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_and_b32 s2, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s21, 8
+; GFX11-NEXT:    s_and_b32 s4, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s23, 8
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_or_b32 s3, s5, s4
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v26
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v27
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v14
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    s_and_b32 s3, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_or_b32_e32 v2, v29, v2
+; GFX11-NEXT:    v_or_b32_e32 v3, v30, v3
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v23
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_or_b32_e32 v5, v31, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_or_b32_e32 v6, v13, v6
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v12
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_lshl_b32 s6, s29, 8
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v2, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    v_or_b32_e32 v0, v32, v0
+; GFX11-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v25
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v16
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v18
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v24
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_or_b32_e32 v1, v28, v1
+; GFX11-NEXT:    v_or_b32_e32 v5, v11, v5
+; GFX11-NEXT:    v_or_b32_e32 v8, v15, v8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_or_b32_e32 v9, v17, v9
+; GFX11-NEXT:    v_or_b32_e32 v10, v19, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_or_b32_e32 v0, v22, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v11
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_mov_b32_e32 v2, s2
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v10
+; GFX11-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:  .LBB35_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB35_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
+; GFX11-NEXT:    s_branch .LBB35_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -18541,488 +17546,237 @@ define inreg <40 x i8> @bitcast_v20i16_to_v40i8_scalar(<20 x i16> inreg %a, i32
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v20i16_to_v40i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s17, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s16, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s19, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s21, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s18, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[9:10], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[13:14]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 24, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 8, v13
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_5
-; GFX11-TRUE16-NEXT:  .LBB49_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s63
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s61
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s59
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s57
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s47
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s4
-; GFX11-TRUE16-NEXT:  .LBB49_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v37, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v39, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v32, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v29, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v14, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v30, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v24, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v19, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v10, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v36, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v3, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v4, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v6
-; GFX11-TRUE16-NEXT:    s_clause 0x2
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[11:14], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v20i16_to_v40i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v6, s17, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v5, s16, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v10, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v9, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v14, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v4, s19, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v2, s21, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v1, s20, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v3, s18, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 24, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 24, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 8, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 8, v13
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_5
-; GFX11-FAKE16-NEXT:  .LBB49_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v4, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s20 :: v_dual_mov_b32 v2, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v48, s63 :: v_dual_mov_b32 v39, s61
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s62 :: v_dual_mov_b32 v37, s60
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s59 :: v_dual_mov_b32 v35, s58
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s56 :: v_dual_mov_b32 v33, s57
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s47 :: v_dual_mov_b32 v31, s46
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s45 :: v_dual_mov_b32 v29, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s44 :: v_dual_mov_b32 v27, s42
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v25, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s28 :: v_dual_mov_b32 v23, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s27 :: v_dual_mov_b32 v21, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s25 :: v_dual_mov_b32 v19, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v7, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s15 :: v_dual_mov_b32 v15, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s6 :: v_dual_mov_b32 v17, s4
-; GFX11-FAKE16-NEXT:  .LBB49_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xff, v34
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v39, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v34, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v32, v31
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v29, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v30
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v9, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v10, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v27
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v24, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v22
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v37, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v19, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v3, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v4, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v17
-; GFX11-FAKE16-NEXT:    s_clause 0x2
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v20i16_to_v40i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s22, 0
+; GFX11-NEXT:    s_mov_b32 s14, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s15, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s22, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s24, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s23, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s26, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s27, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s29, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s28, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s41, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s47, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s59, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_4
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v6, s17, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v5, s16, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v10, s3, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v9, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v14, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v4, s19, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v2, s21, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v1, s20, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v3, s18, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b64 v[16:17], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[17:18], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 24, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 24, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 8, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 8, v13
+; GFX11-NEXT:    s_branch .LBB49_5
+; GFX11-NEXT:  .LBB49_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    s_branch .LBB49_2
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT:    v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
+; GFX11-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
+; GFX11-NEXT:    v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v4, s19
+; GFX11-NEXT:    v_dual_mov_b32 v1, s20 :: v_dual_mov_b32 v2, s21
+; GFX11-NEXT:    v_dual_mov_b32 v48, s63 :: v_dual_mov_b32 v39, s61
+; GFX11-NEXT:    v_dual_mov_b32 v38, s62 :: v_dual_mov_b32 v37, s60
+; GFX11-NEXT:    v_dual_mov_b32 v36, s59 :: v_dual_mov_b32 v35, s58
+; GFX11-NEXT:    v_dual_mov_b32 v34, s56 :: v_dual_mov_b32 v33, s57
+; GFX11-NEXT:    v_dual_mov_b32 v32, s47 :: v_dual_mov_b32 v31, s46
+; GFX11-NEXT:    v_dual_mov_b32 v30, s45 :: v_dual_mov_b32 v29, s43
+; GFX11-NEXT:    v_dual_mov_b32 v28, s44 :: v_dual_mov_b32 v27, s42
+; GFX11-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v25, s40
+; GFX11-NEXT:    v_dual_mov_b32 v24, s28 :: v_dual_mov_b32 v23, s29
+; GFX11-NEXT:    v_dual_mov_b32 v22, s27 :: v_dual_mov_b32 v21, s26
+; GFX11-NEXT:    v_dual_mov_b32 v20, s25 :: v_dual_mov_b32 v19, s23
+; GFX11-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v7, s12
+; GFX11-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s10
+; GFX11-NEXT:    v_dual_mov_b32 v8, s15 :: v_dual_mov_b32 v15, s8
+; GFX11-NEXT:    v_dual_mov_b32 v16, s6 :: v_dual_mov_b32 v17, s4
+; GFX11-NEXT:  .LBB49_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
+; GFX11-NEXT:    v_and_b32_e32 v39, 0xff, v39
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xff, v34
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v32
+; GFX11-NEXT:    v_lshlrev_b32_e32 v31, 8, v31
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v48
+; GFX11-NEXT:    v_or_b32_e32 v17, v39, v17
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v35
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
+; GFX11-NEXT:    v_or_b32_e32 v16, v34, v16
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v33
+; GFX11-NEXT:    v_or_b32_e32 v31, v32, v31
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
+; GFX11-NEXT:    v_or_b32_e32 v15, v29, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 16, v31
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v30
+; GFX11-NEXT:    v_lshlrev_b32_e32 v30, 16, v15
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v17
+; GFX11-NEXT:    v_or_b32_e32 v15, v9, v16
+; GFX11-NEXT:    v_or_b32_e32 v16, v10, v29
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v28
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v26
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xff, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v9
+; GFX11-NEXT:    v_or_b32_e32 v9, v10, v17
+; GFX11-NEXT:    v_or_b32_e32 v10, v24, v11
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xff, v22
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v21
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v38
+; GFX11-NEXT:    v_or_b32_e32 v36, v37, v36
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v25
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v23
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v17
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v20
+; GFX11-NEXT:    v_or_b32_e32 v7, v19, v7
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v18
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v8
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v35
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v30
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v9
+; GFX11-NEXT:    v_or_b32_e32 v7, v3, v10
+; GFX11-NEXT:    v_or_b32_e32 v8, v4, v11
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v12
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v17
+; GFX11-NEXT:    s_clause 0x2
+; GFX11-NEXT:    scratch_store_b128 v0, v[13:16], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -21005,434 +19759,212 @@ define inreg <20 x i16> @bitcast_v40i8_to_v20i16_scalar(<40 x i8> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_branch .LBB51_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v40i8_to_v20i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v15 :: v_dual_mov_b32 v25, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v14 :: v_dual_mov_b32 v27, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v36, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, v10 :: v_dual_mov_b32 v24, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v23, v7 :: v_dual_mov_b32 v26, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v5 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v3 :: v_dual_mov_b32 v38, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB51_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v37
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v33
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v38
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v1, 0xffff, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v3, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v6, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v10, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v7, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v12, 16, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB51_3
-; GFX11-TRUE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s25, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s17, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s19, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s10
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s9, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v18
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v31
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v32
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v9, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v10, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v12, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v10, 16, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v5, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v11, 16, v7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v9, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:  .LBB51_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB51_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-TRUE16-NEXT:    s_branch .LBB51_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v40i8_to_v20i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, v14 :: v_dual_mov_b32 v28, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, v10 :: v_dual_mov_b32 v26, v8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, v4 :: v_dual_mov_b32 v29, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v21
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB51_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v23
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e64 v2, 0xffff, s10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v30
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v24
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v5, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v6, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v6, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v0, 16, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v9, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v2, 16, v1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB51_3
-; GFX11-FAKE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v27
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v36, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 3, v26
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s17, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s10, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s19, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v7
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s6, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s9, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 3, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v35, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v33, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 3, v23
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v18
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v25
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v34, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v30
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v19, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v37, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v32, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v22, v4
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v21, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v17, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v31, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e64 v10, 0xffff, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v3, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v1, 16, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:  .LBB51_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB51_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-FAKE16-NEXT:    s_branch .LBB51_2
+; GFX11-LABEL: bitcast_v40i8_to_v20i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
+; GFX11-NEXT:    v_dual_mov_b32 v25, v14 :: v_dual_mov_b32 v28, v12
+; GFX11-NEXT:    v_dual_mov_b32 v27, v10 :: v_dual_mov_b32 v26, v8
+; GFX11-NEXT:    v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v0
+; GFX11-NEXT:    v_dual_mov_b32 v30, v4 :: v_dual_mov_b32 v29, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v31, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 8, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 8, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v21
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB51_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v23
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    s_and_b32 s11, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s12, s29, 8
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_or_b32 s10, s11, s12
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v22
+; GFX11-NEXT:    v_and_b32_e64 v2, 0xffff, s10
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v30
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v24
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v28
+; GFX11-NEXT:    v_lshl_or_b32 v4, v0, 16, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v26
+; GFX11-NEXT:    v_or_b32_e32 v2, v3, v31
+; GFX11-NEXT:    v_or_b32_e32 v3, v5, v34
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v29
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v27
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v33
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v25
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v18
+; GFX11-NEXT:    v_or_b32_e32 v9, v6, v35
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v16
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v36
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v37
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v20
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v19
+; GFX11-NEXT:    v_or_b32_e32 v12, v6, v17
+; GFX11-NEXT:    v_lshl_or_b32 v6, v0, 16, v3
+; GFX11-NEXT:    v_mov_b32_e32 v0, s5
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v32
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v7
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v21
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v7, v9, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v8, v12, 16, v13
+; GFX11-NEXT:    v_mov_b32_e32 v3, s8
+; GFX11-NEXT:    v_lshl_or_b32 v9, v10, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v5, v2, 16, v1
+; GFX11-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB51_3
+; GFX11-NEXT:  .LBB51_2: ; %cmp.true
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v27
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_lshl_b32 s5, s29, 8
+; GFX11-NEXT:    s_and_b32 s4, s28, 0xff
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    s_and_b32 s5, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    s_and_b32 s6, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s27, 8
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_or_b32 s6, s7, s6
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v28
+; GFX11-NEXT:    v_or_b32_e32 v4, v36, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v26
+; GFX11-NEXT:    s_or_b32 s7, s8, s7
+; GFX11-NEXT:    s_and_b32 s8, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s23, 8
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_or_b32 s8, s9, s8
+; GFX11-NEXT:    s_and_b32 s9, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s17, 8
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_or_b32 s9, s10, s9
+; GFX11-NEXT:    s_and_b32 s10, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s19, 8
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x300, v4
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v7
+; GFX11-NEXT:    s_or_b32 s10, s11, s10
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-NEXT:    s_addk_i32 s9, 0x300
+; GFX11-NEXT:    s_addk_i32 s10, 0x300
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v29
+; GFX11-NEXT:    v_or_b32_e32 v5, v35, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v4, v33, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v23
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v18
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v25
+; GFX11-NEXT:    s_addk_i32 s7, 0x300
+; GFX11-NEXT:    s_addk_i32 s8, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v20
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x300, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, v34, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v30
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v16
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_or_b32_e32 v0, v19, v0
+; GFX11-NEXT:    v_or_b32_e32 v2, v37, v2
+; GFX11-NEXT:    v_or_b32_e32 v7, v32, v7
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v4, v22, v4
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_or_b32_e32 v1, v21, v1
+; GFX11-NEXT:    v_or_b32_e32 v3, v17, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-NEXT:    v_or_b32_e32 v6, v31, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-NEXT:    v_and_b32_e64 v10, 0xffff, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v4, v4, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v5, v6, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v7, v8, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v8, v3, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v9, v1, 16, v0
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:  .LBB51_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB51_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-NEXT:    s_branch .LBB51_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -25141,488 +23673,237 @@ define inreg <40 x i8> @bitcast_v20f16_to_v40i8_scalar(<20 x half> inreg %a, i32
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v20f16_to_v40i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB61_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB61_4
-; GFX11-TRUE16-NEXT:  .LBB61_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s19 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s21 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s18 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[9:10], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[13:14]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 24, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 8, v13
-; GFX11-TRUE16-NEXT:    s_branch .LBB61_5
-; GFX11-TRUE16-NEXT:  .LBB61_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB61_2
-; GFX11-TRUE16-NEXT:  .LBB61_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s63
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s61
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s59
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s57
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s47
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s4
-; GFX11-TRUE16-NEXT:  .LBB61_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, v37, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v39, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v32, v31
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v29, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v14, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v30, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v24, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v19, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v10, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v36, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v3, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v4, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v6
-; GFX11-TRUE16-NEXT:    s_clause 0x2
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[11:14], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v20f16_to_v40i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB61_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB61_4
-; GFX11-FAKE16-NEXT:  .LBB61_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v6, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v5, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v10, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v9, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v14, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v4, 0x200, s19 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, 0x200, s21 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v1, 0x200, s20 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v3, 0x200, s18 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 24, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 24, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 8, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 8, v13
-; GFX11-FAKE16-NEXT:    s_branch .LBB61_5
-; GFX11-FAKE16-NEXT:  .LBB61_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    s_branch .LBB61_2
-; GFX11-FAKE16-NEXT:  .LBB61_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v4, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s20 :: v_dual_mov_b32 v2, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v48, s63 :: v_dual_mov_b32 v39, s61
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s62 :: v_dual_mov_b32 v37, s60
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s59 :: v_dual_mov_b32 v35, s58
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s56 :: v_dual_mov_b32 v33, s57
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s47 :: v_dual_mov_b32 v31, s46
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s45 :: v_dual_mov_b32 v29, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s44 :: v_dual_mov_b32 v27, s42
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v25, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s28 :: v_dual_mov_b32 v23, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s27 :: v_dual_mov_b32 v21, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s25 :: v_dual_mov_b32 v19, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v7, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s15 :: v_dual_mov_b32 v15, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s6 :: v_dual_mov_b32 v17, s4
-; GFX11-FAKE16-NEXT:  .LBB61_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xff, v34
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v48
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v39, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v34, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v32, v31
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v29, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v30
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v9, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v10, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v27
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v24, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v22
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v38
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v37, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v19, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v35
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v3, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v4, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v17
-; GFX11-FAKE16-NEXT:    s_clause 0x2
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v20f16_to_v40i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s22, 0
+; GFX11-NEXT:    s_mov_b32 s14, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB61_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s15, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s22, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s24, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s23, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s26, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s27, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s29, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s28, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s41, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s47, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s59, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
+; GFX11-NEXT:    s_cbranch_vccnz .LBB61_4
+; GFX11-NEXT:  .LBB61_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s19 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s21 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s20 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s18 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b64 v[16:17], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[17:18], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 24, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 24, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 8, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 8, v13
+; GFX11-NEXT:    s_branch .LBB61_5
+; GFX11-NEXT:  .LBB61_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    s_branch .LBB61_2
+; GFX11-NEXT:  .LBB61_4:
+; GFX11-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT:    v_dual_mov_b32 v9, s2 :: v_dual_mov_b32 v10, s3
+; GFX11-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
+; GFX11-NEXT:    v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v4, s19
+; GFX11-NEXT:    v_dual_mov_b32 v1, s20 :: v_dual_mov_b32 v2, s21
+; GFX11-NEXT:    v_dual_mov_b32 v48, s63 :: v_dual_mov_b32 v39, s61
+; GFX11-NEXT:    v_dual_mov_b32 v38, s62 :: v_dual_mov_b32 v37, s60
+; GFX11-NEXT:    v_dual_mov_b32 v36, s59 :: v_dual_mov_b32 v35, s58
+; GFX11-NEXT:    v_dual_mov_b32 v34, s56 :: v_dual_mov_b32 v33, s57
+; GFX11-NEXT:    v_dual_mov_b32 v32, s47 :: v_dual_mov_b32 v31, s46
+; GFX11-NEXT:    v_dual_mov_b32 v30, s45 :: v_dual_mov_b32 v29, s43
+; GFX11-NEXT:    v_dual_mov_b32 v28, s44 :: v_dual_mov_b32 v27, s42
+; GFX11-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v25, s40
+; GFX11-NEXT:    v_dual_mov_b32 v24, s28 :: v_dual_mov_b32 v23, s29
+; GFX11-NEXT:    v_dual_mov_b32 v22, s27 :: v_dual_mov_b32 v21, s26
+; GFX11-NEXT:    v_dual_mov_b32 v20, s25 :: v_dual_mov_b32 v19, s23
+; GFX11-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v7, s12
+; GFX11-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s10
+; GFX11-NEXT:    v_dual_mov_b32 v8, s15 :: v_dual_mov_b32 v15, s8
+; GFX11-NEXT:    v_dual_mov_b32 v16, s6 :: v_dual_mov_b32 v17, s4
+; GFX11-NEXT:  .LBB61_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 8, v48
+; GFX11-NEXT:    v_and_b32_e32 v39, 0xff, v39
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 8, v35
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xff, v34
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v33
+; GFX11-NEXT:    v_and_b32_e32 v32, 0xff, v32
+; GFX11-NEXT:    v_lshlrev_b32_e32 v31, 8, v31
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v48
+; GFX11-NEXT:    v_or_b32_e32 v17, v39, v17
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v35
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
+; GFX11-NEXT:    v_or_b32_e32 v16, v34, v16
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v33
+; GFX11-NEXT:    v_or_b32_e32 v31, v32, v31
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
+; GFX11-NEXT:    v_or_b32_e32 v15, v29, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 16, v31
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v30
+; GFX11-NEXT:    v_lshlrev_b32_e32 v30, 16, v15
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v17
+; GFX11-NEXT:    v_or_b32_e32 v15, v9, v16
+; GFX11-NEXT:    v_or_b32_e32 v16, v10, v29
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 8, v28
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v26
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v38, 8, v38
+; GFX11-NEXT:    v_and_b32_e32 v37, 0xff, v37
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 8, v36
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v9
+; GFX11-NEXT:    v_or_b32_e32 v9, v10, v17
+; GFX11-NEXT:    v_or_b32_e32 v10, v24, v11
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xff, v22
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v21
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v38
+; GFX11-NEXT:    v_or_b32_e32 v36, v37, v36
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v25
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v23
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v17
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v20
+; GFX11-NEXT:    v_or_b32_e32 v7, v19, v7
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v18
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v12, 16, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v8
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v35
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v30
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v9
+; GFX11-NEXT:    v_or_b32_e32 v7, v3, v10
+; GFX11-NEXT:    v_or_b32_e32 v8, v4, v11
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v12
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v17
+; GFX11-NEXT:    s_clause 0x2
+; GFX11-NEXT:    scratch_store_b128 v0, v[13:16], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -27539,434 +25820,212 @@ define inreg <20 x half> @bitcast_v40i8_to_v20f16_scalar(<40 x i8> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_branch .LBB63_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v40i8_to_v20f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v15 :: v_dual_mov_b32 v25, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v14 :: v_dual_mov_b32 v27, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v29, v11 :: v_dual_mov_b32 v36, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, v10 :: v_dual_mov_b32 v24, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v23, v7 :: v_dual_mov_b32 v26, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v5 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v3 :: v_dual_mov_b32 v38, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB63_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v37
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v33
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v38
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v1, 0xffff, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v3, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v6, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v27
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v10, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v7, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v12, 16, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB63_3
-; GFX11-TRUE16-NEXT:  .LBB63_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s25, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s17, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s19, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s10
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s9, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v18
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v31
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v23
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v38
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v32
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v9, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v10, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v12, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v10, 16, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v5, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v11, 16, v7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v9, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:  .LBB63_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB63_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-TRUE16-NEXT:    s_branch .LBB63_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v40i8_to_v20f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, v14 :: v_dual_mov_b32 v28, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, v10 :: v_dual_mov_b32 v26, v8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, v4 :: v_dual_mov_b32 v29, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v21
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB63_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v23
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e64 v2, 0xffff, s10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v30
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v24
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v31
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v5, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v6, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v36
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v6, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v0, 16, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v9, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v2, 16, v1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB63_3
-; GFX11-FAKE16-NEXT:  .LBB63_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v27
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v36, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 3, v26
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s23, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s17, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s10, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s19, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v24
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v7
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s6, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s9, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 3, v29
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v35, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v33, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 3, v23
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v18
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v25
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v34, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v30
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v19, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v37, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v32, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v22, v4
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v21, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v17, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v31, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e64 v10, 0xffff, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v3, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v1, 16, v0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:  .LBB63_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB63_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-FAKE16-NEXT:    s_branch .LBB63_2
+; GFX11-LABEL: bitcast_v40i8_to_v20f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
+; GFX11-NEXT:    v_dual_mov_b32 v25, v14 :: v_dual_mov_b32 v28, v12
+; GFX11-NEXT:    v_dual_mov_b32 v27, v10 :: v_dual_mov_b32 v26, v8
+; GFX11-NEXT:    v_dual_mov_b32 v24, v6 :: v_dual_mov_b32 v23, v0
+; GFX11-NEXT:    v_dual_mov_b32 v30, v4 :: v_dual_mov_b32 v29, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v31, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 8, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 8, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v21
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB63_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v23
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    s_and_b32 s11, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s12, s29, 8
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_or_b32 s10, s11, s12
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v22
+; GFX11-NEXT:    v_and_b32_e64 v2, 0xffff, s10
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v30
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v24
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v28
+; GFX11-NEXT:    v_lshl_or_b32 v4, v0, 16, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v26
+; GFX11-NEXT:    v_or_b32_e32 v2, v3, v31
+; GFX11-NEXT:    v_or_b32_e32 v3, v5, v34
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v29
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v27
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v33
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v25
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v18
+; GFX11-NEXT:    v_or_b32_e32 v9, v6, v35
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v16
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v36
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v37
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v20
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v19
+; GFX11-NEXT:    v_or_b32_e32 v12, v6, v17
+; GFX11-NEXT:    v_lshl_or_b32 v6, v0, 16, v3
+; GFX11-NEXT:    v_mov_b32_e32 v0, s5
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v32
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v7
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v21
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v7, v9, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v8, v12, 16, v13
+; GFX11-NEXT:    v_mov_b32_e32 v3, s8
+; GFX11-NEXT:    v_lshl_or_b32 v9, v10, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v5, v2, 16, v1
+; GFX11-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB63_3
+; GFX11-NEXT:  .LBB63_2: ; %cmp.true
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v27
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_lshl_b32 s5, s29, 8
+; GFX11-NEXT:    s_and_b32 s4, s28, 0xff
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    s_and_b32 s5, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    s_and_b32 s6, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s27, 8
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_or_b32 s6, s7, s6
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v28
+; GFX11-NEXT:    v_or_b32_e32 v4, v36, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v26
+; GFX11-NEXT:    s_or_b32 s7, s8, s7
+; GFX11-NEXT:    s_and_b32 s8, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s23, 8
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_or_b32 s8, s9, s8
+; GFX11-NEXT:    s_and_b32 s9, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s17, 8
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_or_b32 s9, s10, s9
+; GFX11-NEXT:    s_and_b32 s10, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s19, 8
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v24
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x300, v4
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v7
+; GFX11-NEXT:    s_or_b32 s10, s11, s10
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-NEXT:    s_addk_i32 s9, 0x300
+; GFX11-NEXT:    s_addk_i32 s10, 0x300
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v29
+; GFX11-NEXT:    v_or_b32_e32 v5, v35, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v4, v33, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v23
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v18
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v25
+; GFX11-NEXT:    s_addk_i32 s7, 0x300
+; GFX11-NEXT:    s_addk_i32 s8, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v20
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x300, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, v34, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v30
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v16
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_or_b32_e32 v0, v19, v0
+; GFX11-NEXT:    v_or_b32_e32 v2, v37, v2
+; GFX11-NEXT:    v_or_b32_e32 v7, v32, v7
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v4, v22, v4
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    v_or_b32_e32 v1, v21, v1
+; GFX11-NEXT:    v_or_b32_e32 v3, v17, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-NEXT:    v_or_b32_e32 v6, v31, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
+; GFX11-NEXT:    v_and_b32_e64 v10, 0xffff, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v4, v4, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v5, v6, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v7, v8, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v8, v3, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v9, v1, 16, v0
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:  .LBB63_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB63_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-NEXT:    s_branch .LBB63_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -31945,490 +30004,241 @@ define inreg <5 x double> @bitcast_v40i8_to_v5f64_scalar(<40 x i8> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_branch .LBB73_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v40i8_to_v5f64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v15 :: v_dual_mov_b32 v23, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v14 :: v_dual_mov_b32 v25, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v11 :: v_dual_mov_b32 v27, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v29, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v8 :: v_dual_mov_b32 v31, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v5 :: v_dual_mov_b32 v33, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v4 :: v_dual_mov_b32 v35, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB73_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v36
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v33
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v30
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v12, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB73_3
-; GFX11-TRUE16-NEXT:  .LBB73_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v37
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v36
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v34
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v32
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v6, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v11, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v13, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v14, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v15, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v12, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB73_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB73_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-TRUE16-NEXT:    s_branch .LBB73_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v40i8_to_v5f64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, v14 :: v_dual_mov_b32 v29, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v27, v8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, v6 :: v_dual_mov_b32 v24, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, v2 :: v_dual_mov_b32 v26, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v21
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB73_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v26
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v27
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v37
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v30
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v31
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v24
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB73_3
-; GFX11-FAKE16-NEXT:  .LBB73_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v23
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v27
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v30
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s24, 0xff
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v34, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v35, v3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v26
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v31, v6
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v28
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v37, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v36, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v24
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 3, v16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 3, v18
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v29
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 3, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v33, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v22, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v17, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v19, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v32, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v21, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s2
-; GFX11-FAKE16-NEXT:  .LBB73_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB73_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-FAKE16-NEXT:    s_branch .LBB73_2
+; GFX11-LABEL: bitcast_v40i8_to_v5f64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
+; GFX11-NEXT:    v_dual_mov_b32 v30, v14 :: v_dual_mov_b32 v29, v12
+; GFX11-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v27, v8
+; GFX11-NEXT:    v_dual_mov_b32 v23, v6 :: v_dual_mov_b32 v24, v4
+; GFX11-NEXT:    v_dual_mov_b32 v25, v2 :: v_dual_mov_b32 v26, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 8, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v31, 8, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v21
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB73_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v26
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v23
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v27
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v37
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v28
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v34
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v35
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v30
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_and_b32 s8, s8, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s9, s9, 16
+; GFX11-NEXT:    s_and_b32 s10, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s29, 8
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v36
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v31
+; GFX11-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v24
+; GFX11-NEXT:    s_and_b32 s10, s10, 0xffff
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v29
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v16
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v20
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v2, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s8
+; GFX11-NEXT:    v_or_b32_e32 v4, s10, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v25
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v33
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v22
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v17
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v19
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v32
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v11
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-NEXT:    v_mov_b32_e32 v0, s5
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v10
+; GFX11-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB73_3
+; GFX11-NEXT:  .LBB73_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s17, 8
+; GFX11-NEXT:    s_and_b32 s3, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_and_b32 s2, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s21, 8
+; GFX11-NEXT:    s_and_b32 s4, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s23, 8
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_or_b32 s3, s5, s4
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v23
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v27
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v30
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    s_and_b32 s3, s24, 0xff
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    s_lshl_b32 s4, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    v_or_b32_e32 v2, v34, v2
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v3, v35, v3
+; GFX11-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v26
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_or_b32_e32 v6, v31, v6
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v28
+; GFX11-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_lshl_b32 s6, s29, 8
+; GFX11-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v2, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    v_or_b32_e32 v0, v37, v0
+; GFX11-NEXT:    v_or_b32_e32 v5, v36, v5
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v24
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v16
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v18
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v29
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v20
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v25
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_or_b32_e32 v1, v33, v1
+; GFX11-NEXT:    v_or_b32_e32 v5, v22, v5
+; GFX11-NEXT:    v_or_b32_e32 v8, v17, v8
+; GFX11-NEXT:    v_or_b32_e32 v9, v19, v9
+; GFX11-NEXT:    v_or_b32_e32 v0, v32, v0
+; GFX11-NEXT:    v_or_b32_e32 v10, v21, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v11
+; GFX11-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v10
+; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s2
+; GFX11-NEXT:  .LBB73_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB73_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-NEXT:    s_branch .LBB73_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -34129,486 +31939,231 @@ define inreg <40 x i8> @bitcast_v5f64_to_v40i8_scalar(<5 x double> inreg %a, i32
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v5f64_to_v40i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s45, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB75_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s45
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB75_4
-; GFX11-TRUE16-NEXT:  .LBB75_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[12:13], s[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[10:11], s[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[7:8], s[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[4:5], s[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[1:2], s[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s1, v13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s3, v11
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s17, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s19, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s21, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[14:15], 24, v[10:11]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[2:3], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[4:5]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[12:13]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 8, v12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 8
-; GFX11-TRUE16-NEXT:    s_branch .LBB75_5
-; GFX11-TRUE16-NEXT:  .LBB75_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB75_2
-; GFX11-TRUE16-NEXT:  .LBB75_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s63
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s61
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s59
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s57
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s47
-; GFX11-TRUE16-NEXT:  .LBB75_5: ; %end
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s44
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s42
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v14
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v17, v14
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s41
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s29
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v10, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v16
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v19, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_and_b32 v10, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s26
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v10, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v15
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s25
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v4, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v5
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s23
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s24
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v5
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s14
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s0
-; GFX11-TRUE16-NEXT:    s_clause 0x2
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[12:15], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[4:7], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v5f64_to_v40i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s45, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB75_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s45
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB75_4
-; GFX11-FAKE16-NEXT:  .LBB75_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[12:13], s[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[10:11], s[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[7:8], s[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[4:5], s[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[1:2], s[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s1, v13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s3, v11
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s17, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s19, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s21, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[14:15], 24, v[10:11]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[2:3], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[4:5]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[7:8]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[12:13]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 8, v12
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s1, 8
-; GFX11-FAKE16-NEXT:    s_branch .LBB75_5
-; GFX11-FAKE16-NEXT:  .LBB75_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    s_branch .LBB75_2
-; GFX11-FAKE16-NEXT:  .LBB75_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s0 :: v_dual_mov_b32 v7, s16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s2 :: v_dual_mov_b32 v1, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s18 :: v_dual_mov_b32 v15, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s6 :: v_dual_mov_b32 v5, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v19, s63
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s12 :: v_dual_mov_b32 v17, s61
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s62 :: v_dual_mov_b32 v13, s59
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s60 :: v_dual_mov_b32 v9, s57
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s58 :: v_dual_mov_b32 v11, s56
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s47 :: v_dual_mov_b32 v6, s46
-; GFX11-FAKE16-NEXT:  .LBB75_5: ; %end
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s1, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s44, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s43, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s42, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s41, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s3, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s29, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s40, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v19, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v14
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v13, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v13, s0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s26, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s23, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v15
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v15, s1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s28, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v9, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s27, 0xff
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s25, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v11
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s19, 0xff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v6
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s24, 0xff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s21, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s22, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s15, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s14, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v10, v14
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v4, v5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, s1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s0
-; GFX11-FAKE16-NEXT:    s_clause 0x2
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[12:15], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v5f64_to_v40i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s22, 0
+; GFX11-NEXT:    s_mov_b32 s45, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB75_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s14, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s15, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s22, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s47, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s23, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s24, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s57, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s26, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s27, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s28, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s59, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s29, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s40, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s63, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s45
+; GFX11-NEXT:    s_cbranch_vccnz .LBB75_4
+; GFX11-NEXT:  .LBB75_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[12:13], s[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[10:11], s[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[7:8], s[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[4:5], s[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[1:2], s[20:21], 1.0
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v13
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v11
+; GFX11-NEXT:    v_readfirstlane_b32 s17, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s19, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s21, v2
+; GFX11-NEXT:    v_lshrrev_b64 v[14:15], 24, v[10:11]
+; GFX11-NEXT:    v_lshrrev_b64 v[2:3], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[5:6], 24, v[4:5]
+; GFX11-NEXT:    v_lshrrev_b64 v[8:9], 24, v[7:8]
+; GFX11-NEXT:    v_lshrrev_b64 v[15:16], 24, v[12:13]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 8, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 8, v12
+; GFX11-NEXT:    s_lshr_b32 s14, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s15, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s22, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s23, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s24, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s25, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s26, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s27, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s28, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s29, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s40, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s1, 8
+; GFX11-NEXT:    s_branch .LBB75_5
+; GFX11-NEXT:  .LBB75_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    s_branch .LBB75_2
+; GFX11-NEXT:  .LBB75_4:
+; GFX11-NEXT:    v_dual_mov_b32 v12, s0 :: v_dual_mov_b32 v7, s16
+; GFX11-NEXT:    v_dual_mov_b32 v10, s2 :: v_dual_mov_b32 v1, s20
+; GFX11-NEXT:    v_dual_mov_b32 v4, s18 :: v_dual_mov_b32 v15, s4
+; GFX11-NEXT:    v_dual_mov_b32 v14, s6 :: v_dual_mov_b32 v5, s10
+; GFX11-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v19, s63
+; GFX11-NEXT:    v_dual_mov_b32 v2, s12 :: v_dual_mov_b32 v17, s61
+; GFX11-NEXT:    v_dual_mov_b32 v20, s62 :: v_dual_mov_b32 v13, s59
+; GFX11-NEXT:    v_dual_mov_b32 v18, s60 :: v_dual_mov_b32 v9, s57
+; GFX11-NEXT:    v_dual_mov_b32 v16, s58 :: v_dual_mov_b32 v11, s56
+; GFX11-NEXT:    v_dual_mov_b32 v3, s47 :: v_dual_mov_b32 v6, s46
+; GFX11-NEXT:  .LBB75_5: ; %end
+; GFX11-NEXT:    s_and_b32 s0, s1, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s44, 8
+; GFX11-NEXT:    s_and_b32 s2, s43, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s42, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s4
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_lshl_b32 s2, s41, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s3, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s29, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_and_b32 s2, s40, 0xff
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
+; GFX11-NEXT:    v_or_b32_e32 v15, v19, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 8, v14
+; GFX11-NEXT:    v_or_b32_e32 v8, v13, v8
+; GFX11-NEXT:    v_mov_b32_e32 v13, s0
+; GFX11-NEXT:    v_or_b32_e32 v12, v12, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-NEXT:    s_and_b32 s0, s17, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s26, 8
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_or_b32_e32 v12, v12, v15
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v17
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
+; GFX11-NEXT:    v_or_b32_e32 v14, v15, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v16
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
+; GFX11-NEXT:    s_lshl_b32 s3, s23, 8
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v18
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v15
+; GFX11-NEXT:    v_mov_b32_e32 v15, s1
+; GFX11-NEXT:    s_lshl_b32 s1, s28, 8
+; GFX11-NEXT:    v_or_b32_e32 v5, v9, v5
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s27, 0xff
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_lshl_b32 s2, s25, 8
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 8, v11
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s19, 0xff
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v6
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_and_b32 s2, s24, 0xff
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_or_b32_e32 v2, v3, v2
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_mov_b32_e32 v8, s0
+; GFX11-NEXT:    s_and_b32 s0, s21, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s22, 8
+; GFX11-NEXT:    s_and_b32 s3, s15, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s14, 8
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-NEXT:    s_or_b32 s0, s0, s2
+; GFX11-NEXT:    s_or_b32 s2, s3, s4
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_or_b32_e32 v14, v10, v14
+; GFX11-NEXT:    s_or_b32 s0, s0, s2
+; GFX11-NEXT:    v_or_b32_e32 v9, v4, v5
+; GFX11-NEXT:    v_mov_b32_e32 v10, s1
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX11-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-NEXT:    s_clause 0x2
+; GFX11-NEXT:    scratch_store_b128 v0, v[12:15], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[7:10], off offset:16
+; GFX11-NEXT:    scratch_store_b64 v0, v[1:2], off offset:32
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -36580,490 +34135,241 @@ define inreg <5 x i64> @bitcast_v40i8_to_v5i64_scalar(<40 x i8> inreg %a, i32 in
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
 ; GFX9-NEXT:    s_branch .LBB77_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v40i8_to_v5i64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v15 :: v_dual_mov_b32 v23, v13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v14 :: v_dual_mov_b32 v25, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v11 :: v_dual_mov_b32 v27, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v29, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v8 :: v_dual_mov_b32 v31, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v5 :: v_dual_mov_b32 v33, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v4 :: v_dual_mov_b32 v35, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB77_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v37
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v36
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v33
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v30
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v12, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB77_3
-; GFX11-TRUE16-NEXT:  .LBB77_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
-; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v37
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v36
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v34
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v32
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s6, s5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v6, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v7, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v24
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v11, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v13, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v14, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v15, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v12, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:  .LBB77_3: ; %end
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB77_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-TRUE16-NEXT:    s_branch .LBB77_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v40i8_to_v5i64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, v14 :: v_dual_mov_b32 v29, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v27, v8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, v6 :: v_dual_mov_b32 v24, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, v2 :: v_dual_mov_b32 v26, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 8, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v21
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB77_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s7, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v26
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v27
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s25, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v37
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v30
-; GFX11-FAKE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s8, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v31
-; GFX11-FAKE16-NEXT:    s_or_b32 s10, s10, s11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v24
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v25
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v33
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB77_3
-; GFX11-FAKE16-NEXT:  .LBB77_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s1, s0
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s3, s2
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s19, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s4, s3
-; GFX11-FAKE16-NEXT:    s_addk_i32 s1, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s21, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s23, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-FAKE16-NEXT:    s_addk_i32 s2, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v23
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v27
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v30
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s24, 0xff
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s25, 8
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v34, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v35, v3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s27, 8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v26
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v31, v6
-; GFX11-FAKE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v28
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s29, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v2, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v37, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v36, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v24
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 3, v16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 3, v18
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v29
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 3, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, s5, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v33, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v22, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v17, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v19, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v32, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v21, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v0, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s2
-; GFX11-FAKE16-NEXT:  .LBB77_3: ; %end
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB77_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
-; GFX11-FAKE16-NEXT:    s_branch .LBB77_2
+; GFX11-LABEL: bitcast_v40i8_to_v5i64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v22
+; GFX11-NEXT:    v_dual_mov_b32 v30, v14 :: v_dual_mov_b32 v29, v12
+; GFX11-NEXT:    v_dual_mov_b32 v28, v10 :: v_dual_mov_b32 v27, v8
+; GFX11-NEXT:    v_dual_mov_b32 v23, v6 :: v_dual_mov_b32 v24, v4
+; GFX11-NEXT:    v_dual_mov_b32 v25, v2 :: v_dual_mov_b32 v26, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v37, 8, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 8, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v34, 8, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 8, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 8, v11
+; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 8, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v31, 8, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v21
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB77_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_and_b32 s5, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s1, 8
+; GFX11-NEXT:    s_and_b32 s7, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s3, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s7, s7, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s8, s8, 16
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v26
+; GFX11-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v23
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v27
+; GFX11-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v37
+; GFX11-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s27, 8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v28
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v34
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v35
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v30
+; GFX11-NEXT:    s_or_b32 s9, s9, s10
+; GFX11-NEXT:    s_and_b32 s8, s8, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s9, s9, 16
+; GFX11-NEXT:    s_and_b32 s10, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s29, 8
+; GFX11-NEXT:    s_or_b32 s8, s8, s9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v36
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_or_b32_e32 v6, v6, v31
+; GFX11-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v24
+; GFX11-NEXT:    s_and_b32 s10, s10, 0xffff
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v29
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v16
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v18
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v20
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v2, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s8
+; GFX11-NEXT:    v_or_b32_e32 v4, s10, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v25
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v33
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v22
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v17
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v19
+; GFX11-NEXT:    v_or_b32_e32 v0, v0, v32
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v11
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-NEXT:    v_mov_b32_e32 v0, s5
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v10
+; GFX11-NEXT:    v_dual_mov_b32 v1, s6 :: v_dual_mov_b32 v2, s7
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB77_3
+; GFX11-NEXT:  .LBB77_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-NEXT:    s_or_b32 s0, s1, s0
+; GFX11-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-NEXT:    s_addk_i32 s0, 0x300
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s17, 8
+; GFX11-NEXT:    s_and_b32 s3, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-NEXT:    s_or_b32 s1, s2, s1
+; GFX11-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_and_b32 s2, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s21, 8
+; GFX11-NEXT:    s_and_b32 s4, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s23, 8
+; GFX11-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-NEXT:    s_or_b32 s3, s5, s4
+; GFX11-NEXT:    s_addk_i32 s2, 0x300
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v23
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v27
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v30
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    s_and_b32 s3, s24, 0xff
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    s_lshl_b32 s4, s25, 8
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    v_or_b32_e32 v2, v34, v2
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-NEXT:    v_or_b32_e32 v3, v35, v3
+; GFX11-NEXT:    s_or_b32 s3, s4, s3
+; GFX11-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v26
+; GFX11-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-NEXT:    v_or_b32_e32 v6, v31, v6
+; GFX11-NEXT:    s_addk_i32 s3, 0x300
+; GFX11-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v28
+; GFX11-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_lshl_b32 s6, s29, 8
+; GFX11-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v2, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    v_or_b32_e32 v0, v37, v0
+; GFX11-NEXT:    v_or_b32_e32 v5, v36, v5
+; GFX11-NEXT:    s_or_b32 s5, s6, s5
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v24
+; GFX11-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v16
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v18
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v29
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v20
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v25
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xff, v0
+; GFX11-NEXT:    v_or_b32_e32 v1, v33, v1
+; GFX11-NEXT:    v_or_b32_e32 v5, v22, v5
+; GFX11-NEXT:    v_or_b32_e32 v8, v17, v8
+; GFX11-NEXT:    v_or_b32_e32 v9, v19, v9
+; GFX11-NEXT:    v_or_b32_e32 v0, v32, v0
+; GFX11-NEXT:    v_or_b32_e32 v10, v21, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v5
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_or_b32_e32 v7, v7, v11
+; GFX11-NEXT:    v_or_b32_e32 v5, v0, v1
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v10
+; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s2
+; GFX11-NEXT:  .LBB77_3: ; %end
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB77_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+; GFX11-NEXT:    s_branch .LBB77_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -38799,471 +36105,220 @@ define inreg <40 x i8> @bitcast_v5i64_to_v40i8_scalar(<5 x i64> inreg %a, i32 in
 ; GFX9-NEXT:    ; implicit-def: $sgpr26
 ; GFX9-NEXT:    s_branch .LBB79_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v5i64_to_v40i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB79_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s63
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB79_3
-; GFX11-TRUE16-NEXT:  .LBB79_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s29, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB79_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s62
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s61
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s12
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s60
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s59
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s58
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s57
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s47
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s46
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s45
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s10, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s44
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s40
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s25
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s14
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-TRUE16-NEXT:    s_clause 0x2
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b64 v0, v[9:10], off offset:32
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB79_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr22_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB79_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v5i64_to_v40i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s22, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s63, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB79_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s63
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB79_3
-; GFX11-FAKE16-NEXT:  .LBB79_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s28, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s29, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB79_3: ; %end
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s62, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s61, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s12, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s60, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s59, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s58, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s57, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s56, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s10, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s47, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s46, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s45, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s44, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s43, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s8, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s42, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s41, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s40, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s29, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s28, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s19, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s27, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s25, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s24, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s23, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s4, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s21, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s22, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s15, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s14, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-FAKE16-NEXT:    s_clause 0x2
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b64 v0, v[9:10], off offset:32
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB79_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr22
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    s_branch .LBB79_2
+; GFX11-LABEL: bitcast_v5i64_to_v40i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s22, 0
+; GFX11-NEXT:    s_mov_b32 s63, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB79_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s14, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s15, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s22, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s23, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s24, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s25, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s26, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s27, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s28, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s29, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s41, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s59, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s63
+; GFX11-NEXT:    s_cbranch_vccnz .LBB79_3
+; GFX11-NEXT:  .LBB79_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s14, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s15, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s22, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s23, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s24, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s25, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s26, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s27, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s28, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s29, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s41, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s59, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s0, 8
+; GFX11-NEXT:  .LBB79_3: ; %end
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s62, 8
+; GFX11-NEXT:    s_and_b32 s7, s61, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s12, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s5
+; GFX11-NEXT:    s_or_b32 s5, s7, s9
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s60, 8
+; GFX11-NEXT:    s_and_b32 s9, s59, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s58, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s5
+; GFX11-NEXT:    s_or_b32 s1, s1, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s57, 8
+; GFX11-NEXT:    s_and_b32 s7, s56, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s10, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s5
+; GFX11-NEXT:    s_or_b32 s5, s7, s9
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s47, 8
+; GFX11-NEXT:    s_and_b32 s9, s46, 0xff
+; GFX11-NEXT:    s_lshl_b32 s10, s45, 8
+; GFX11-NEXT:    s_or_b32 s3, s3, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s10
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s5
+; GFX11-NEXT:    s_or_b32 s3, s3, s7
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    s_and_b32 s0, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s44, 8
+; GFX11-NEXT:    s_and_b32 s2, s43, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s8, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s17, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s42, 8
+; GFX11-NEXT:    s_and_b32 s5, s41, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s40, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s29, 8
+; GFX11-NEXT:    s_and_b32 s5, s28, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s6
+; GFX11-NEXT:    s_and_b32 s5, s19, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s27, 8
+; GFX11-NEXT:    s_and_b32 s7, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s25, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s6
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    s_and_b32 s0, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s24, 8
+; GFX11-NEXT:    s_and_b32 s2, s23, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s4, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s21, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s22, 8
+; GFX11-NEXT:    s_and_b32 s4, s15, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s14, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
+; GFX11-NEXT:    s_clause 0x2
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-NEXT:    scratch_store_b64 v0, v[9:10], off offset:32
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB79_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr29
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr27
+; GFX11-NEXT:    ; implicit-def: $sgpr26
+; GFX11-NEXT:    ; implicit-def: $sgpr25
+; GFX11-NEXT:    ; implicit-def: $sgpr24
+; GFX11-NEXT:    ; implicit-def: $sgpr23
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr22
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    s_branch .LBB79_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll
index 7b756bce857bc..bfed1f4304dd9 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll
@@ -2083,64 +2083,34 @@ define inreg <4 x i8> @bitcast_i32_to_v4i8_scalar(i32 inreg %a, i32 inreg %b) {
 ; GFX9-NEXT:    ; implicit-def: $sgpr6
 ; GFX9-NEXT:    s_branch .LBB21_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_i32_to_v4i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB21_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s1, s0, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB21_3
-; GFX11-TRUE16-NEXT:  .LBB21_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s1, s0, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB21_3: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB21_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr1_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB21_2
-;
-; GFX11-FAKE16-LABEL: bitcast_i32_to_v4i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB21_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s1, s0, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB21_3
-; GFX11-FAKE16-NEXT:  .LBB21_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s1, s0, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB21_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB21_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr1
-; GFX11-FAKE16-NEXT:    s_branch .LBB21_2
+; GFX11-LABEL: bitcast_i32_to_v4i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB21_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s1, s0, 24
+; GFX11-NEXT:    s_lshr_b32 s2, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s3, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB21_3
+; GFX11-NEXT:  .LBB21_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_lshr_b32 s2, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s1, s0, 24
+; GFX11-NEXT:    s_lshr_b32 s3, s0, 8
+; GFX11-NEXT:  .LBB21_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s3
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB21_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr1
+; GFX11-NEXT:    s_branch .LBB21_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -4360,65 +4330,34 @@ define inreg <4 x i8> @bitcast_f32_to_v4i8_scalar(float inreg %a, i32 inreg %b)
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s6
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_f32_to_v4i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB41_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s0, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB41_4
-; GFX11-TRUE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, s0, 1.0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB41_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB41_2
-; GFX11-TRUE16-NEXT:  .LBB41_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s3
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_f32_to_v4i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB41_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s0, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB41_4
-; GFX11-FAKE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, s0, 1.0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB41_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    s_branch .LBB41_2
-; GFX11-FAKE16-NEXT:  .LBB41_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_f32_to_v4i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB41_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s3, s0, 24
+; GFX11-NEXT:    s_lshr_b32 s2, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
+; GFX11-NEXT:    s_cbranch_vccnz .LBB41_4
+; GFX11-NEXT:  .LBB41_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f32_e64 v0, s0, 1.0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB41_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    s_branch .LBB41_2
+; GFX11-NEXT:  .LBB41_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s4
+; GFX11-NEXT:    v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -6347,65 +6286,34 @@ define inreg <4 x i8> @bitcast_v2i16_to_v4i8_scalar(<2 x i16> inreg %a, i32 inre
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s6
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v2i16_to_v4i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB57_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s0, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB57_4
-; GFX11-TRUE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB57_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB57_2
-; GFX11-TRUE16-NEXT:  .LBB57_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s3
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v2i16_to_v4i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB57_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s0, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB57_4
-; GFX11-FAKE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB57_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    s_branch .LBB57_2
-; GFX11-FAKE16-NEXT:  .LBB57_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v2i16_to_v4i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB57_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s3, s0, 24
+; GFX11-NEXT:    s_lshr_b32 s2, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
+; GFX11-NEXT:    s_cbranch_vccnz .LBB57_4
+; GFX11-NEXT:  .LBB57_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB57_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    s_branch .LBB57_2
+; GFX11-NEXT:  .LBB57_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s4
+; GFX11-NEXT:    v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -8029,65 +7937,34 @@ define inreg <4 x i8> @bitcast_v2f16_to_v4i8_scalar(<2 x half> inreg %a, i32 inr
 ; GFX9-NEXT:    v_mov_b32_e32 v2, s6
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v2f16_to_v4i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB69_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s0, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB69_4
-; GFX11-TRUE16-NEXT:  .LBB69_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB69_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB69_2
-; GFX11-TRUE16-NEXT:  .LBB69_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s3
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v2f16_to_v4i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB69_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s0, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB69_4
-; GFX11-FAKE16-NEXT:  .LBB69_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB69_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    s_branch .LBB69_2
-; GFX11-FAKE16-NEXT:  .LBB69_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v2f16_to_v4i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB69_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s3, s0, 24
+; GFX11-NEXT:    s_lshr_b32 s2, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
+; GFX11-NEXT:    s_cbranch_vccnz .LBB69_4
+; GFX11-NEXT:  .LBB69_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB69_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    s_branch .LBB69_2
+; GFX11-NEXT:  .LBB69_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s4
+; GFX11-NEXT:    v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -9398,111 +9275,57 @@ define inreg <4 x i8> @bitcast_v2bf16_to_v4i8_scalar(<2 x bfloat> inreg %a, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s7
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v2bf16_to_v4i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB77_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s0, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB77_4
-; GFX11-TRUE16-NEXT:  .LBB77_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB77_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB77_2
-; GFX11-TRUE16-NEXT:  .LBB77_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s3
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v2bf16_to_v4i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB77_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s0, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB77_4
-; GFX11-FAKE16-NEXT:  .LBB77_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB77_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    s_branch .LBB77_2
-; GFX11-FAKE16-NEXT:  .LBB77_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v1, s3
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v2bf16_to_v4i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB77_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s2, s0, 24
+; GFX11-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s3, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s1
+; GFX11-NEXT:    s_cbranch_vccnz .LBB77_4
+; GFX11-NEXT:  .LBB77_2: ; %cmp.true
+; GFX11-NEXT:    s_lshl_b32 s1, s0, 16
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s1
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v3, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 24, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB77_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    s_branch .LBB77_2
+; GFX11-NEXT:  .LBB77_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s2
+; GFX11-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v1, s3
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -10186,64 +10009,34 @@ define inreg <4 x i8> @bitcast_v1i32_to_v4i8_scalar(<1 x i32> inreg %a, i32 inre
 ; GFX9-NEXT:    ; implicit-def: $sgpr6
 ; GFX9-NEXT:    s_branch .LBB81_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v1i32_to_v4i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB81_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s1, s0, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB81_3
-; GFX11-TRUE16-NEXT:  .LBB81_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s1, s0, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB81_3: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB81_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr1_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB81_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v1i32_to_v4i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB81_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s1, s0, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB81_3
-; GFX11-FAKE16-NEXT:  .LBB81_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s1, s0, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB81_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB81_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr1
-; GFX11-FAKE16-NEXT:    s_branch .LBB81_2
+; GFX11-LABEL: bitcast_v1i32_to_v4i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB81_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s1, s0, 24
+; GFX11-NEXT:    s_lshr_b32 s2, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s3, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB81_3
+; GFX11-NEXT:  .LBB81_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_lshr_b32 s2, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s1, s0, 24
+; GFX11-NEXT:    s_lshr_b32 s3, s0, 8
+; GFX11-NEXT:  .LBB81_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s3
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB81_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr1
+; GFX11-NEXT:    s_branch .LBB81_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll
index 6ec9c1177c180..d774e561c6398 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll
@@ -369,114 +369,59 @@ define inreg <3 x half> @bitcast_v3bf16_to_v3f16_scalar(<3 x bfloat> inreg %a, i
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v3bf16_to_v3f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB1_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB1_4
-; GFX11-TRUE16-NEXT:  .LBB1_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v3, v7 :: v_dual_add_nc_u32 v5, v5, v2
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v4, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v5, v8 :: v_dual_and_b32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, 0x7fc0, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v2, 16, v0
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB1_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB1_2
-; GFX11-TRUE16-NEXT:  .LBB1_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v3bf16_to_v3f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB1_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB1_4
-; GFX11-FAKE16-NEXT:  .LBB1_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v5, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v5, v8 :: v_dual_and_b32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, 0x7fc0, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v2, 16, v0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB1_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB1_2
-; GFX11-FAKE16-NEXT:  .LBB1_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v3bf16_to_v3f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s2, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB1_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
+; GFX11-NEXT:    s_cbranch_vccnz .LBB1_4
+; GFX11-NEXT:  .LBB1_2: ; %cmp.true
+; GFX11-NEXT:    s_lshl_b32 s2, s0, 16
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
+; GFX11-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v3, v0, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v5, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v4, v6, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v5, v8 :: v_dual_and_b32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshl_or_b32 v1, 0x7fc0, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v0, v2, 16, v0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB1_3:
+; GFX11-NEXT:    s_branch .LBB1_2
+; GFX11-NEXT:  .LBB1_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -1074,107 +1019,56 @@ define inreg <3 x i16> @bitcast_v3bf16_to_v3i16_scalar(<3 x bfloat> inreg %a, i3
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v3bf16_to_v3i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB5_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB5_4
-; GFX11-TRUE16-NEXT:  .LBB5_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v3, v7 :: v_dual_add_nc_u32 v5, v5, v2
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v4, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v5, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, 0x7fc0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_or_b32 v0, 0xffff0000, v2, v0
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB5_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB5_2
-; GFX11-TRUE16-NEXT:  .LBB5_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v3bf16_to_v3i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB5_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB5_4
-; GFX11-FAKE16-NEXT:  .LBB5_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v5, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v5, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, 0x7fc0, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_and_or_b32 v0, 0xffff0000, v2, v0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB5_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB5_2
-; GFX11-FAKE16-NEXT:  .LBB5_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v3bf16_to_v3i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s2, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB5_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
+; GFX11-NEXT:    s_cbranch_vccnz .LBB5_4
+; GFX11-NEXT:  .LBB5_2: ; %cmp.true
+; GFX11-NEXT:    s_lshl_b32 s2, s0, 16
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
+; GFX11-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v3, v0, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_bfe_u32 v5, v2, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v2
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, v5, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v4, v6, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_cndmask_b32_e32 v2, v5, v8, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v1, 0x7fc0, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT:    v_and_or_b32 v0, 0xffff0000, v2, v0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB5_3:
+; GFX11-NEXT:    s_branch .LBB5_2
+; GFX11-NEXT:  .LBB5_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
index 1024c2a7f066a..d52451418e49a 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
@@ -10577,781 +10577,362 @@ define inreg <64 x i8> @bitcast_v16i32_to_v64i8_scalar(<16 x i32> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $sgpr56
 ; GFX9-NEXT:    s_branch .LBB25_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v16i32_to_v64i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v17, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s30, 0
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 vcc_lo, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s49, 9
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB25_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s30, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB25_3
-; GFX11-TRUE16-NEXT:  .LBB25_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s30, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB25_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s49
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s48
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s40
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s39
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s38
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s37
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s36
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s35
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s31
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s30
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s95
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s94
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s14
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s93
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s92
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s91
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s90
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s89
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s12
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s88
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s79
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s78
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s76
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s75
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s73
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s72
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s63
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s62
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s61
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s60
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s59
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s58
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s57
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s47
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s46
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s45
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s44
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s42
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v17, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v17, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v17, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v17, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v17, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v17, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v17, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v17, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v17, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v17, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v17, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB25_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr95_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB25_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v16i32_to_v64i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v17, s32 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s30, 0
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 vcc_lo, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s48, 8
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB25_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s30, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s31, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB25_3
-; GFX11-FAKE16-NEXT:  .LBB25_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s30, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s31, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB25_3: ; %end
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s48, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s39, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s40, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s38, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s37, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s36, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s35, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s34, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s28, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s31, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s30, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, vcc_hi, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s95, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s94, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s14, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s93, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s92, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s91, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s90, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s89, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s12, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s19, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s88, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s79, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s78, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s77, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s76, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s10, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s21, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s75, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s74, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s73, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s72, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s63, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s8, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s23, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s62, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s61, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s60, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s59, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s58, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s6, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s25, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s57, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s56, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s47, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s46, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s45, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s27, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s44, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s43, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s42, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v17, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v17, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v17, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v17, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v17, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v17, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v17, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v17, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v17, 0
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v17, off, s32 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB25_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr95
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    s_branch .LBB25_2
+; GFX11-LABEL: bitcast_v16i32_to_v64i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
+; GFX11-NEXT:    scratch_store_b32 off, v17, s32 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v17, s30, 0
+; GFX11-NEXT:    s_cmp_lg_u32 s28, 0
+; GFX11-NEXT:    s_mov_b32 vcc_lo, 0
+; GFX11-NEXT:    v_writelane_b32 v17, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v17, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v17, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v17, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v17, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v17, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v17, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v17, s48, 8
+; GFX11-NEXT:    s_cbranch_scc0 .LBB25_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s47, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s60, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s61, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s73, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s74, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s79, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s89, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s91, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s92, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s94, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s95, s16, 8
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s30, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s31, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s34, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s35, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s36, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s37, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s38, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
+; GFX11-NEXT:    s_cbranch_vccnz .LBB25_3
+; GFX11-NEXT:  .LBB25_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s47, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s60, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s61, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s73, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s74, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s79, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s89, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s91, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s92, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s94, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s95, s16, 8
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s30, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s31, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s34, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s35, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s36, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s37, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s38, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s0, 8
+; GFX11-NEXT:  .LBB25_3: ; %end
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s48, 8
+; GFX11-NEXT:    s_and_b32 s7, s39, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s40, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s5
+; GFX11-NEXT:    s_or_b32 s5, s7, s9
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s38, 8
+; GFX11-NEXT:    s_and_b32 s9, s37, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s36, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s5
+; GFX11-NEXT:    s_or_b32 s1, s1, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s35, 8
+; GFX11-NEXT:    s_and_b32 s7, s34, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s28, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s5
+; GFX11-NEXT:    s_or_b32 s5, s7, s9
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s31, 8
+; GFX11-NEXT:    s_and_b32 s9, s30, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, vcc_hi, 8
+; GFX11-NEXT:    s_or_b32 s3, s3, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s5
+; GFX11-NEXT:    s_or_b32 s3, s3, s7
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    s_and_b32 s0, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s95, 8
+; GFX11-NEXT:    s_and_b32 s2, s94, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s14, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s17, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s93, 8
+; GFX11-NEXT:    s_and_b32 s5, s92, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s91, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s90, 8
+; GFX11-NEXT:    s_and_b32 s5, s89, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s12, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s5, s19, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s88, 8
+; GFX11-NEXT:    s_and_b32 s9, s79, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s78, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    s_and_b32 s0, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s77, 8
+; GFX11-NEXT:    s_and_b32 s2, s76, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s10, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s21, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s75, 8
+; GFX11-NEXT:    s_and_b32 s5, s74, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s73, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s72, 8
+; GFX11-NEXT:    s_and_b32 s5, s63, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s8, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s5, s23, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s62, 8
+; GFX11-NEXT:    s_and_b32 s8, s61, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s60, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s7
+; GFX11-NEXT:    s_or_b32 s7, s8, s9
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
+; GFX11-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
+; GFX11-NEXT:    s_and_b32 s0, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s59, 8
+; GFX11-NEXT:    s_and_b32 s2, s58, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s6, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s25, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s57, 8
+; GFX11-NEXT:    s_and_b32 s5, s56, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s47, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s6
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s46, 8
+; GFX11-NEXT:    s_and_b32 s5, s45, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s4
+; GFX11-NEXT:    s_and_b32 s4, s27, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s44, 8
+; GFX11-NEXT:    s_and_b32 s6, s43, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s42, 8
+; GFX11-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-NEXT:    s_or_b32 s5, s6, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s4, s4, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT:    v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[13:16], off offset:48
+; GFX11-NEXT:    v_readlane_b32 s48, v17, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v17, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v17, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v17, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v17, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v17, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v17, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v17, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v17, 0
+; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT:    scratch_load_b32 v17, off, s32 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB25_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr31
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $sgpr95
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    s_branch .LBB25_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -15135,34 +14716,61 @@ define inreg <16 x i32> @bitcast_v64i8_to_v16i32_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-LABEL: bitcast_v64i8_to_v16i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v1 :: v_dual_mov_b32 v54, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, v15 :: v_dual_mov_b32 v32, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v13 :: v_dual_mov_b32 v34, v11
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v12 :: v_dual_mov_b32 v36, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v9 :: v_dual_mov_b32 v38, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v8 :: v_dual_mov_b32 v48, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v5 :: v_dual_mov_b32 v50, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v4 :: v_dual_mov_b32 v52, v2
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v0, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v2, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)
+; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(12)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(11)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 8, v6
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(10)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v8
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(9)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v10
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(8)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v12
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v14
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(6)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v86
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB27_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
@@ -15199,104 +14807,82 @@ define inreg <16 x i32> @bitcast_v64i8_to_v16i32_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-NEXT:    s_and_b32 s8, s9, 0xffff
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v31
 ; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v32
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v38
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v83
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v85
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s28, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v5
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v37
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v68
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v69
+; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
 ; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v8, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v19
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v37
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v6
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v13, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v15, v86
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v71
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v80
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v81
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v65
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v67
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v52
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v7
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v64
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v3
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v30
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v30
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v84
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v15, v86
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v13, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v68
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v96, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v98, 0xff, v65
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v99, 8, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v14, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v86, v87
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v96, v97
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v98, v99
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v51
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v29
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v54
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v55
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v21
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v86, v25
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -15329,204 +14915,179 @@ define inreg <16 x i32> @bitcast_v64i8_to_v16i32_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v54
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v31
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v32
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v38
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v49
+; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v83, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v84, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v85, v2
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v82, v6
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v68, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v69, v11
+; GFX11-TRUE16-NEXT:    s_and_b32 s6, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s29, 8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s6
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v7, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v34
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v16
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v52
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v70, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v71, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v80, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v81, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v65, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v66, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v67, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v7, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v8, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v24
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v8, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v9, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v25
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v3, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v29
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v3
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v83
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v64, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v26
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v28
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v30
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v51
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v49
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v48
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v39
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v30
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v12, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v14, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v69
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v12, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v13, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v80
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v70
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v65
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v66
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v17, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v19, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v20, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v21, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v27, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v29, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v54, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v55, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v17, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v19, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v21, v14
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v23, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v25, v16
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v11
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v2, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v1, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v14, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v14
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v16
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
@@ -25905,811 +25466,391 @@ define inreg <64 x i8> @bitcast_v16f32_to_v64i8_scalar(<16 x float> inreg %a, i3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v16f32_to_v64i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s50, 10
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s94, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s30, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, s17, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, s16, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, s3, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, s2, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v26, s1, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, s0, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, s19, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, s21, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, s23, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, s25, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, s27, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, s26, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, s24, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, s22, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, s20, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, s18, 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[17:18]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[28:29], 24, v[21:22]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[9:10]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[23:24], 24, v[13:14]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[29:30], 24, v[25:26]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 24, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 24, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 24, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_5
-; GFX11-TRUE16-NEXT:  .LBB49_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr95_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v96.l, s50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, s49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v86.l, s48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.l, s39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.l, s38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v83.l, s37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v82.l, s36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v81.l, s35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v80.l, s34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v71.l, s31
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v70.l, s30
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v69.l, s95
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v68.l, s94
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, s93
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, s92
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, s91
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, s90
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, s89
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, s88
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s79
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s78
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s77
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s76
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s75
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s74
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s73
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s72
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s63
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s61
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s59
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s57
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s47
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s4
-; GFX11-TRUE16-NEXT:  .LBB49_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v87, 0xff, v87
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v86
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v85, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v87, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v80, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v81
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v26, v86
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, v85, v84
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v83
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v82, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v22, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v71, v80, v71
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v17, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v69, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v67, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v26, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v21, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v22, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v17, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, v18, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v64
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v55
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v53
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v18, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v23, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v49
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 8, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v23, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v38, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v68, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, v14, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v9, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v10, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v7, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v33
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v18, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v18, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v12, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v7, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v11, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v12, v6
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[25:28], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[66:69], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v40, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v40, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v40, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v40, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v40, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v40, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v40, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v40, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v40, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v16f32_to_v64i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s42, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s49, 9
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s30, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s31, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, s19, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, s18, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, s17, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v15, s16, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, s3, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, s2, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, s1, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, s0, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, s21, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, s23, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, s25, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, s27, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, s26, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, s24, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, s22, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, s20, 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[25:26], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[15:16]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[19:20]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[21:22], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[28:29], 24, v[23:24]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 24, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 24, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 8, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 24, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v23
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_5
-; GFX11-FAKE16-NEXT:  .LBB49_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr95
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s0 :: v_dual_mov_b32 v24, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s2 :: v_dual_mov_b32 v20, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s16 :: v_dual_mov_b32 v16, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s18 :: v_dual_mov_b32 v14, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s22 :: v_dual_mov_b32 v6, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s24 :: v_dual_mov_b32 v4, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s26 :: v_dual_mov_b32 v2, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v96, s49 :: v_dual_mov_b32 v87, s39
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s48 :: v_dual_mov_b32 v85, s38
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v84, s37 :: v_dual_mov_b32 v83, s36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, s34 :: v_dual_mov_b32 v81, s35
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v80, s31 :: v_dual_mov_b32 v71, s30
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, vcc_hi :: v_dual_mov_b32 v69, s94
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v68, s95 :: v_dual_mov_b32 v67, s93
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, s92 :: v_dual_mov_b32 v65, s91
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v64, s89 :: v_dual_mov_b32 v55, s90
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, s88 :: v_dual_mov_b32 v53, s79
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s78 :: v_dual_mov_b32 v51, s76
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, s77 :: v_dual_mov_b32 v49, s75
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v48, s74 :: v_dual_mov_b32 v39, s73
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s63 :: v_dual_mov_b32 v37, s72
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s62 :: v_dual_mov_b32 v35, s61
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s60 :: v_dual_mov_b32 v33, s58
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s59 :: v_dual_mov_b32 v31, s57
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s56 :: v_dual_mov_b32 v29, s47
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v7, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s46 :: v_dual_mov_b32 v11, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s44 :: v_dual_mov_b32 v17, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s43 :: v_dual_mov_b32 v21, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, s10 :: v_dual_mov_b32 v26, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s6 :: v_dual_mov_b32 v28, s4
-; GFX11-FAKE16-NEXT:  .LBB49_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v96
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v87, 0xff, v87
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v85, 0xff, v85
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v23, v96
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v87, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v87, 0xff, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v86
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v83
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v71
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, v23, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v27
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v85, v84
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xff, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v27
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v80
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v67, 0xff, v67
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v86
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v19, v83
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v87, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v71
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v70
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v69, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v16, v68
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v67, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v83, v24, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v84, v19, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v85, v20, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v15, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v16, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v65
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v64
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v55
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v54
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v16, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v25, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v51
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v49
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v39
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v19, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v25, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v28, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v13, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v14, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v9, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v10, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v5, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v33
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v9, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v17, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v31
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v22
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v17, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v11, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v12, v8
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[82:85], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[23:26], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v40, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v40, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v40, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v40, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v40, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v40, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v40, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v40, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v16f32_to_v64i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    s_cmp_lg_u32 s28, 0
+; GFX11-NEXT:    s_mov_b32 s42, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v40, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v40, s49, 9
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s44, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s57, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s74, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s75, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s79, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s88, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s89, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s92, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s93, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s95, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s94, s16, 16
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s30, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s31, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s35, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s34, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s36, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s37, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s38, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s49, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_4
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f32_e64 v14, s19, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v13, s18, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v16, s17, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v15, s16, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v20, s3, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v19, s2, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v24, s1, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v23, s0, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v10, s21, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v6, s23, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v4, s25, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v2, s27, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v1, s26, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v3, s24, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v5, s22, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v9, s20, 1.0
+; GFX11-NEXT:    v_lshrrev_b64 v[25:26], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b64 v[26:27], 24, v[15:16]
+; GFX11-NEXT:    v_lshrrev_b64 v[27:28], 24, v[19:20]
+; GFX11-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[17:18], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b64 v[21:22], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[28:29], 24, v[23:24]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 24, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 24, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 8, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 8, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 24, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 8, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 24, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 8, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 8, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 24, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 8, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 8, v23
+; GFX11-NEXT:    s_branch .LBB49_5
+; GFX11-NEXT:  .LBB49_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr31
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr95
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    s_branch .LBB49_2
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    v_dual_mov_b32 v23, s0 :: v_dual_mov_b32 v24, s1
+; GFX11-NEXT:    v_dual_mov_b32 v19, s2 :: v_dual_mov_b32 v20, s3
+; GFX11-NEXT:    v_dual_mov_b32 v15, s16 :: v_dual_mov_b32 v16, s17
+; GFX11-NEXT:    v_dual_mov_b32 v13, s18 :: v_dual_mov_b32 v14, s19
+; GFX11-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
+; GFX11-NEXT:    v_dual_mov_b32 v5, s22 :: v_dual_mov_b32 v6, s23
+; GFX11-NEXT:    v_dual_mov_b32 v3, s24 :: v_dual_mov_b32 v4, s25
+; GFX11-NEXT:    v_dual_mov_b32 v1, s26 :: v_dual_mov_b32 v2, s27
+; GFX11-NEXT:    v_dual_mov_b32 v96, s49 :: v_dual_mov_b32 v87, s39
+; GFX11-NEXT:    v_dual_mov_b32 v86, s48 :: v_dual_mov_b32 v85, s38
+; GFX11-NEXT:    v_dual_mov_b32 v84, s37 :: v_dual_mov_b32 v83, s36
+; GFX11-NEXT:    v_dual_mov_b32 v82, s34 :: v_dual_mov_b32 v81, s35
+; GFX11-NEXT:    v_dual_mov_b32 v80, s31 :: v_dual_mov_b32 v71, s30
+; GFX11-NEXT:    v_dual_mov_b32 v70, vcc_hi :: v_dual_mov_b32 v69, s94
+; GFX11-NEXT:    v_dual_mov_b32 v68, s95 :: v_dual_mov_b32 v67, s93
+; GFX11-NEXT:    v_dual_mov_b32 v66, s92 :: v_dual_mov_b32 v65, s91
+; GFX11-NEXT:    v_dual_mov_b32 v64, s89 :: v_dual_mov_b32 v55, s90
+; GFX11-NEXT:    v_dual_mov_b32 v54, s88 :: v_dual_mov_b32 v53, s79
+; GFX11-NEXT:    v_dual_mov_b32 v52, s78 :: v_dual_mov_b32 v51, s76
+; GFX11-NEXT:    v_dual_mov_b32 v50, s77 :: v_dual_mov_b32 v49, s75
+; GFX11-NEXT:    v_dual_mov_b32 v48, s74 :: v_dual_mov_b32 v39, s73
+; GFX11-NEXT:    v_dual_mov_b32 v38, s63 :: v_dual_mov_b32 v37, s72
+; GFX11-NEXT:    v_dual_mov_b32 v36, s62 :: v_dual_mov_b32 v35, s61
+; GFX11-NEXT:    v_dual_mov_b32 v34, s60 :: v_dual_mov_b32 v33, s58
+; GFX11-NEXT:    v_dual_mov_b32 v32, s59 :: v_dual_mov_b32 v31, s57
+; GFX11-NEXT:    v_dual_mov_b32 v30, s56 :: v_dual_mov_b32 v29, s47
+; GFX11-NEXT:    v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v7, s40
+; GFX11-NEXT:    v_dual_mov_b32 v18, s46 :: v_dual_mov_b32 v11, s28
+; GFX11-NEXT:    v_dual_mov_b32 v12, s44 :: v_dual_mov_b32 v17, s14
+; GFX11-NEXT:    v_dual_mov_b32 v8, s43 :: v_dual_mov_b32 v21, s12
+; GFX11-NEXT:    v_dual_mov_b32 v25, s10 :: v_dual_mov_b32 v26, s8
+; GFX11-NEXT:    v_dual_mov_b32 v27, s6 :: v_dual_mov_b32 v28, s4
+; GFX11-NEXT:  .LBB49_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xff, v23
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v96, 8, v96
+; GFX11-NEXT:    v_and_b32_e32 v87, 0xff, v87
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
+; GFX11-NEXT:    v_and_b32_e32 v85, 0xff, v85
+; GFX11-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
+; GFX11-NEXT:    v_or_b32_e32 v23, v23, v96
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-NEXT:    v_or_b32_e32 v28, v87, v28
+; GFX11-NEXT:    v_and_b32_e32 v87, 0xff, v82
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshlrev_b32_e32 v86, 8, v86
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v83, 8, v83
+; GFX11-NEXT:    v_lshlrev_b32_e32 v71, 8, v71
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-NEXT:    v_or_b32_e32 v82, v23, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 8, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v81
+; GFX11-NEXT:    v_or_b32_e32 v28, v85, v84
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xff, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v26
+; GFX11-NEXT:    v_or_b32_e32 v20, v20, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v80
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
+; GFX11-NEXT:    v_and_b32_e32 v67, 0xff, v67
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
+; GFX11-NEXT:    v_or_b32_e32 v24, v24, v86
+; GFX11-NEXT:    v_or_b32_e32 v19, v19, v83
+; GFX11-NEXT:    v_or_b32_e32 v23, v87, v23
+; GFX11-NEXT:    v_or_b32_e32 v28, v28, v71
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v70
+; GFX11-NEXT:    v_or_b32_e32 v26, v69, v26
+; GFX11-NEXT:    v_or_b32_e32 v16, v16, v68
+; GFX11-NEXT:    v_or_b32_e32 v66, v67, v66
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
+; GFX11-NEXT:    v_or_b32_e32 v83, v24, v27
+; GFX11-NEXT:    v_or_b32_e32 v84, v19, v23
+; GFX11-NEXT:    v_or_b32_e32 v85, v20, v28
+; GFX11-NEXT:    v_or_b32_e32 v23, v15, v26
+; GFX11-NEXT:    v_or_b32_e32 v24, v16, v66
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v65
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xff, v64
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v25
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v55
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v54
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v53
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v52
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v15
+; GFX11-NEXT:    v_or_b32_e32 v15, v16, v19
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v20
+; GFX11-NEXT:    v_or_b32_e32 v16, v25, v26
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v27
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v51
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v21
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v50
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v49
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v48
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v39
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_or_b32_e32 v19, v19, v20
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v21
+; GFX11-NEXT:    v_or_b32_e32 v20, v25, v26
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v27
+; GFX11-NEXT:    v_or_b32_e32 v17, v28, v17
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-NEXT:    v_or_b32_e32 v25, v13, v15
+; GFX11-NEXT:    v_or_b32_e32 v26, v14, v16
+; GFX11-NEXT:    v_or_b32_e32 v13, v9, v19
+; GFX11-NEXT:    v_or_b32_e32 v14, v10, v20
+; GFX11-NEXT:    v_or_b32_e32 v15, v5, v17
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 8, v37
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 8, v35
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v34
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v33
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v32
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v9, v10
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v16
+; GFX11-NEXT:    v_or_b32_e32 v9, v17, v11
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v31
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v29
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v22
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v19
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v16
+; GFX11-NEXT:    v_or_b32_e32 v7, v17, v7
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v18
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_or_b32_e32 v16, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v1, v3, v9
+; GFX11-NEXT:    v_or_b32_e32 v2, v4, v10
+; GFX11-NEXT:    v_or_b32_e32 v3, v11, v7
+; GFX11-NEXT:    v_or_b32_e32 v4, v12, v8
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b128 v0, v[82:85], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[23:26], off offset:16
+; GFX11-NEXT:    scratch_store_b128 v0, v[13:16], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
+; GFX11-NEXT:    v_readlane_b32 s49, v40, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v40, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v40, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -30493,34 +29634,61 @@ define inreg <16 x float> @bitcast_v64i8_to_v16f32_scalar(<64 x i8> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v64i8_to_v16f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v1 :: v_dual_mov_b32 v54, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, v15 :: v_dual_mov_b32 v32, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v13 :: v_dual_mov_b32 v34, v11
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v12 :: v_dual_mov_b32 v36, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v9 :: v_dual_mov_b32 v38, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v8 :: v_dual_mov_b32 v48, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v5 :: v_dual_mov_b32 v50, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v4 :: v_dual_mov_b32 v52, v2
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v0, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v2, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)
+; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(12)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(11)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 8, v6
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(10)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v8
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(9)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v10
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(8)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v12
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v14
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(6)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v86
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB51_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
@@ -30557,104 +29725,82 @@ define inreg <16 x float> @bitcast_v64i8_to_v16f32_scalar(<64 x i8> inreg %a, i3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s8, s9, 0xffff
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v31
 ; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v32
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v38
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v83
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v85
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s28, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v5
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v37
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v68
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v69
+; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
 ; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v8, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v19
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v37
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v6
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v13, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v15, v86
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v71
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v80
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v81
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v65
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v67
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v52
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v7
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v64
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v3
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v30
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v30
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v84
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v15, v86
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v13, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v68
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v96, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v98, 0xff, v65
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v99, 8, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v14, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v86, v87
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v96, v97
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v98, v99
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v51
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v29
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v54
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v55
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v21
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v86, v25
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -30687,204 +29833,179 @@ define inreg <16 x float> @bitcast_v64i8_to_v16f32_scalar(<64 x i8> inreg %a, i3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v54
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v31
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v32
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v38
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v49
+; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v83, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v84, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v85, v2
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v82, v6
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v68, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v69, v11
+; GFX11-TRUE16-NEXT:    s_and_b32 s6, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s29, 8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s6
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v7, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v34
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v16
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v52
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v70, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v71, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v80, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v81, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v65, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v66, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v67, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v7, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v8, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v24
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v8, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v9, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v25
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v3, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v29
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v3
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v83
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v64, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v26
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v28
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v30
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v51
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v49
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v48
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v39
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v30
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v12, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v14, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v69
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v12, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v13, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v80
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v70
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v65
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v66
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v17, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v19, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v20, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v21, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v27, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v29, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v54, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v55, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v17, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v19, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v21, v14
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v23, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v25, v16
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v11
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v2, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v1, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v14, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v14
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v16
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
@@ -40564,781 +39685,362 @@ define inreg <64 x i8> @bitcast_v8i64_to_v64i8_scalar(<8 x i64> inreg %a, i32 in
 ; GFX9-NEXT:    ; implicit-def: $sgpr56
 ; GFX9-NEXT:    s_branch .LBB69_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v8i64_to_v64i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v17, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s30, 0
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 vcc_lo, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s49, 9
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB69_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s30, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB69_3
-; GFX11-TRUE16-NEXT:  .LBB69_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s30, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB69_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s49
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s48
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s40
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s39
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s38
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s37
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s36
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s35
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s31
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s30
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s95
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s94
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s14
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s93
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s92
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s91
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s90
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s89
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s12
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s88
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s79
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s78
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s76
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s75
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s73
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s72
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s63
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s62
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s61
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s60
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s59
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s58
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s57
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s47
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s46
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s45
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s44
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s42
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v17, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v17, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v17, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v17, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v17, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v17, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v17, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v17, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v17, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v17, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v17, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB69_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr95_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB69_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v8i64_to_v64i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v17, s32 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s30, 0
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 vcc_lo, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s48, 8
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB69_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s30, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s31, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB69_3
-; GFX11-FAKE16-NEXT:  .LBB69_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s30, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s31, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB69_3: ; %end
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s48, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s39, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s40, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s38, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s37, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s36, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s35, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s34, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s28, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s31, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s30, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, vcc_hi, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s95, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s94, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s14, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s93, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s92, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s91, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s90, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s89, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s12, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s19, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s88, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s79, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s78, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s77, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s76, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s10, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s21, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s75, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s74, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s73, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s72, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s63, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s8, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s23, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s62, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s61, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s60, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s59, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s58, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s6, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s25, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s57, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s56, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s47, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s46, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s45, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s27, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s44, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s43, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s42, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v17, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v17, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v17, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v17, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v17, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v17, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v17, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v17, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v17, 0
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v17, off, s32 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB69_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr95
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    s_branch .LBB69_2
+; GFX11-LABEL: bitcast_v8i64_to_v64i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
+; GFX11-NEXT:    scratch_store_b32 off, v17, s32 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v17, s30, 0
+; GFX11-NEXT:    s_cmp_lg_u32 s28, 0
+; GFX11-NEXT:    s_mov_b32 vcc_lo, 0
+; GFX11-NEXT:    v_writelane_b32 v17, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v17, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v17, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v17, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v17, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v17, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v17, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v17, s48, 8
+; GFX11-NEXT:    s_cbranch_scc0 .LBB69_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s47, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s60, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s61, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s73, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s74, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s79, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s89, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s91, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s92, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s94, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s95, s16, 8
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s30, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s31, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s34, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s35, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s36, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s37, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s38, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
+; GFX11-NEXT:    s_cbranch_vccnz .LBB69_3
+; GFX11-NEXT:  .LBB69_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s47, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s60, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s61, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s73, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s74, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s79, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s89, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s91, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s92, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s94, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s95, s16, 8
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s30, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s31, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s34, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s35, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s36, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s37, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s38, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s0, 8
+; GFX11-NEXT:  .LBB69_3: ; %end
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s48, 8
+; GFX11-NEXT:    s_and_b32 s7, s39, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s40, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s5
+; GFX11-NEXT:    s_or_b32 s5, s7, s9
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s38, 8
+; GFX11-NEXT:    s_and_b32 s9, s37, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s36, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s5
+; GFX11-NEXT:    s_or_b32 s1, s1, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s35, 8
+; GFX11-NEXT:    s_and_b32 s7, s34, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s28, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s5
+; GFX11-NEXT:    s_or_b32 s5, s7, s9
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s31, 8
+; GFX11-NEXT:    s_and_b32 s9, s30, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, vcc_hi, 8
+; GFX11-NEXT:    s_or_b32 s3, s3, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s5
+; GFX11-NEXT:    s_or_b32 s3, s3, s7
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    s_and_b32 s0, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s95, 8
+; GFX11-NEXT:    s_and_b32 s2, s94, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s14, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s17, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s93, 8
+; GFX11-NEXT:    s_and_b32 s5, s92, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s91, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s90, 8
+; GFX11-NEXT:    s_and_b32 s5, s89, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s12, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s5, s19, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s88, 8
+; GFX11-NEXT:    s_and_b32 s9, s79, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s78, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    s_and_b32 s0, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s77, 8
+; GFX11-NEXT:    s_and_b32 s2, s76, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s10, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s21, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s75, 8
+; GFX11-NEXT:    s_and_b32 s5, s74, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s73, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s72, 8
+; GFX11-NEXT:    s_and_b32 s5, s63, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s8, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s5, s23, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s62, 8
+; GFX11-NEXT:    s_and_b32 s8, s61, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s60, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s7
+; GFX11-NEXT:    s_or_b32 s7, s8, s9
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
+; GFX11-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
+; GFX11-NEXT:    s_and_b32 s0, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s59, 8
+; GFX11-NEXT:    s_and_b32 s2, s58, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s6, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s25, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s57, 8
+; GFX11-NEXT:    s_and_b32 s5, s56, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s47, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s6
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s46, 8
+; GFX11-NEXT:    s_and_b32 s5, s45, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s4, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s4
+; GFX11-NEXT:    s_and_b32 s4, s27, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s44, 8
+; GFX11-NEXT:    s_and_b32 s6, s43, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s42, 8
+; GFX11-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-NEXT:    s_or_b32 s5, s6, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s4, s4, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT:    v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[13:16], off offset:48
+; GFX11-NEXT:    v_readlane_b32 s48, v17, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v17, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v17, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v17, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v17, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v17, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v17, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v17, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v17, 0
+; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT:    scratch_load_b32 v17, off, s32 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB69_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr31
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $sgpr95
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    s_branch .LBB69_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -45122,34 +43824,61 @@ define inreg <8 x i64> @bitcast_v64i8_to_v8i64_scalar(<64 x i8> inreg %a, i32 in
 ; GFX11-TRUE16-LABEL: bitcast_v64i8_to_v8i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v1 :: v_dual_mov_b32 v54, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, v15 :: v_dual_mov_b32 v32, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v13 :: v_dual_mov_b32 v34, v11
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v12 :: v_dual_mov_b32 v36, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v9 :: v_dual_mov_b32 v38, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v8 :: v_dual_mov_b32 v48, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v5 :: v_dual_mov_b32 v50, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v4 :: v_dual_mov_b32 v52, v2
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v0, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v2, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)
+; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(12)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(11)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 8, v6
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(10)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v8
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(9)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v10
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(8)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v12
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v14
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(6)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v86
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB71_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
@@ -45186,104 +43915,82 @@ define inreg <8 x i64> @bitcast_v64i8_to_v8i64_scalar(<64 x i8> inreg %a, i32 in
 ; GFX11-TRUE16-NEXT:    s_and_b32 s8, s9, 0xffff
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v31
 ; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v32
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v38
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v83
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v85
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s28, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v5
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v37
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v68
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v69
+; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
 ; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v8, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v19
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v37
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v6
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v13, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v15, v86
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v71
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v80
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v81
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v65
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v67
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v52
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v7
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v64
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v3
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v30
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v30
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v84
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v15, v86
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v13, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v68
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v96, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v98, 0xff, v65
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v99, 8, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v14, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v86, v87
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v96, v97
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v98, v99
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v51
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v29
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v54
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v55
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v21
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v86, v25
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -45316,204 +44023,179 @@ define inreg <8 x i64> @bitcast_v64i8_to_v8i64_scalar(<64 x i8> inreg %a, i32 in
 ; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v54
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v31
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v32
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v38
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v49
+; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v83, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v84, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v85, v2
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v82, v6
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v68, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v69, v11
+; GFX11-TRUE16-NEXT:    s_and_b32 s6, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s29, 8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s6
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v7, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v34
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v16
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v52
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v70, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v71, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v80, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v81, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v65, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v66, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v67, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v7, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v8, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v24
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v8, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v9, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v25
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v3, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v29
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v3
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v83
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v64, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v26
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v28
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v30
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v51
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v49
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v48
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v39
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v30
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v12, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v14, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v69
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v12, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v13, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v80
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v70
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v65
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v66
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v17, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v19, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v20, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v21, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v27, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v29, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v54, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v55, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v17, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v19, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v21, v14
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v23, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v25, v16
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v11
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v2, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v1, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v14, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v14
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v16
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
@@ -54320,808 +53002,381 @@ define inreg <64 x i8> @bitcast_v8f64_to_v64i8_scalar(<8 x double> inreg %a, i32
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v8f64_to_v64i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v33, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s30, 0
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s90, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v33, s50, 10
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB85_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s94, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s30, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB85_4
-; GFX11-TRUE16-NEXT:  .LBB85_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[24:25], s[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[20:21], s[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[16:17], s[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[14:15], s[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[9:10], s[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[7:8], s[22:23], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[4:5], s[24:25], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[1:2], s[26:27], 1.0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s1, v25
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s3, v21
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s17, v17
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s19, v15
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s21, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s23, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s25, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s27, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[20:21]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[2:3], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[4:5]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[12:13], 24, v[9:10]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[18:19], 24, v[14:15]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[22:23], 24, v[16:17]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[24:25]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 8, v24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s1, 8
-; GFX11-TRUE16-NEXT:    s_branch .LBB85_5
-; GFX11-TRUE16-NEXT:  .LBB85_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr95_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB85_2
-; GFX11-TRUE16-NEXT:  .LBB85_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s31
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s95
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s30
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s93
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s94
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s91
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s92
-; GFX11-TRUE16-NEXT:  .LBB85_5: ; %end
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s89
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s88
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s79
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v26
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s78
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v20, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v29, v26
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s76
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s77
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v20, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v22
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v31, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v25, v22
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v25, s0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v24, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s74
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v24, v27
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v27, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s75
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v22
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v21, v18
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s73
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s62
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s72
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s63
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v12
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v14, v18
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v17, v12
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v19, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s61
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s60
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s59
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s56
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s58
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s57
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v10
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v10
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s47
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s46
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s45
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v7, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v1, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v2
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s44
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s42
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v9, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v16, v20
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v4, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s1
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[24:27], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[16:19], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[7:10], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v33, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v33, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v33, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v33, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v33, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v33, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v33, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v33, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v33, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v33, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v33, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v33, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v8f64_to_v64i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v33, s32 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v33, s30, 0
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s90, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v33, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v33, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v33, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v33, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v33, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v33, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v33, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v33, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v33, s49, 9
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB85_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s94, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s31, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s30, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB85_4
-; GFX11-FAKE16-NEXT:  .LBB85_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], s[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[24:25], s[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[22:23], s[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[14:15], s[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[10:11], s[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[8:9], s[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[3:4], s[24:25], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[1:2], s[26:27], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[18:19]
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s1, v25
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s3, v23
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s17, v19
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s19, v15
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s21, v11
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s23, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s25, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s27, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[22:23]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[6:7], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[12:13], 24, v[8:9]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[16:17], 24, v[10:11]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[20:21], 24, v[14:15]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[28:29], 24, v[24:25]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 8, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 8, v24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s1, 8
-; GFX11-FAKE16-NEXT:    s_branch .LBB85_5
-; GFX11-FAKE16-NEXT:  .LBB85_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr95
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    s_branch .LBB85_2
-; GFX11-FAKE16-NEXT:  .LBB85_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v3, s24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v1, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v27, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s18 :: v_dual_mov_b32 v5, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s20 :: v_dual_mov_b32 v31, s49
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v29, s39
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s4 :: v_dual_mov_b32 v23, s37
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s8 :: v_dual_mov_b32 v25, s36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s10 :: v_dual_mov_b32 v19, s35
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s12 :: v_dual_mov_b32 v21, s34
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s14 :: v_dual_mov_b32 v15, s31
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s28 :: v_dual_mov_b32 v17, s30
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s48 :: v_dual_mov_b32 v11, vcc_hi
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s38 :: v_dual_mov_b32 v13, s95
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s94 :: v_dual_mov_b32 v2, s92
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s93 :: v_dual_mov_b32 v4, s91
-; GFX11-FAKE16-NEXT:  .LBB85_5: ; %end
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s1, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s89, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s79, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s88, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s3, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s78, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s77, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s76, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 8, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v31, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v22
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v18, v25
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v25, s1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v32
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s75, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s73, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v29, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v22, v24
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v31, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s62, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v23
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v23, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s17, 0xff
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s74, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s72, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s19, 0xff
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v12
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s63, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v19, v20
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v17
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s0 :: v_dual_lshlrev_b32 v20, 16, v15
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v17, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s21, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s61, 8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s59, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s60, 0xff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v13
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s23, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s58, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s57, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s56, 8
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v26
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v8, v11
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v26
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v21
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v13, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s25, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s47, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s46, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s45, 8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v14, v21
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v1, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v5
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s44, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s27, 0xff
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s43, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s42, 8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v18, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s3, s4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v18, v19
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, s1
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[22:25], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[14:17], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[10:13], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v33, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v33, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v33, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v33, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v33, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v33, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v33, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v33, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v33, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v33, 0
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v33, off, s32 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v8f64_to_v64i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
+; GFX11-NEXT:    scratch_store_b32 off, v33, s32 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v33, s30, 0
+; GFX11-NEXT:    s_cmp_lg_u32 s28, 0
+; GFX11-NEXT:    s_mov_b32 s90, 0
+; GFX11-NEXT:    v_writelane_b32 v33, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v33, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v33, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v33, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v33, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v33, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v33, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v33, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v33, s49, 9
+; GFX11-NEXT:    s_cbranch_scc0 .LBB85_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s92, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s46, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s94, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s57, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s23, 8
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s95, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s59, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s60, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s31, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s30, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s62, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s63, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s35, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s34, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s73, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s74, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s37, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s36, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s77, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s38, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s79, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s88, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s49, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
+; GFX11-NEXT:    s_cbranch_vccnz .LBB85_4
+; GFX11-NEXT:  .LBB85_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[18:19], s[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[24:25], s[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[22:23], s[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[14:15], s[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[10:11], s[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[8:9], s[22:23], 1.0
+; GFX11-NEXT:    v_add_f64 v[3:4], s[24:25], 1.0
+; GFX11-NEXT:    v_add_f64 v[1:2], s[26:27], 1.0
+; GFX11-NEXT:    v_lshrrev_b64 v[26:27], 24, v[18:19]
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v25
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v23
+; GFX11-NEXT:    v_readfirstlane_b32 s17, v19
+; GFX11-NEXT:    v_readfirstlane_b32 s19, v15
+; GFX11-NEXT:    v_readfirstlane_b32 s21, v11
+; GFX11-NEXT:    v_readfirstlane_b32 s23, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s25, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s27, v2
+; GFX11-NEXT:    v_lshrrev_b64 v[5:6], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[27:28], 24, v[22:23]
+; GFX11-NEXT:    v_lshrrev_b64 v[6:7], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[12:13], 24, v[8:9]
+; GFX11-NEXT:    v_lshrrev_b64 v[16:17], 24, v[10:11]
+; GFX11-NEXT:    v_lshrrev_b64 v[20:21], 24, v[14:15]
+; GFX11-NEXT:    v_lshrrev_b64 v[28:29], 24, v[24:25]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 8, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 8, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 8, v24
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s46, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s57, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s59, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s60, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s62, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s63, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s73, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s74, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s77, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s79, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s88, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s1, 8
+; GFX11-NEXT:    s_branch .LBB85_5
+; GFX11-NEXT:  .LBB85_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $sgpr31
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr95
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    s_branch .LBB85_2
+; GFX11-NEXT:  .LBB85_4:
+; GFX11-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v3, s24
+; GFX11-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v1, s26
+; GFX11-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v27, s6
+; GFX11-NEXT:    v_dual_mov_b32 v14, s18 :: v_dual_mov_b32 v5, s40
+; GFX11-NEXT:    v_dual_mov_b32 v10, s20 :: v_dual_mov_b32 v31, s49
+; GFX11-NEXT:    v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v29, s39
+; GFX11-NEXT:    v_dual_mov_b32 v28, s4 :: v_dual_mov_b32 v23, s37
+; GFX11-NEXT:    v_dual_mov_b32 v26, s8 :: v_dual_mov_b32 v25, s36
+; GFX11-NEXT:    v_dual_mov_b32 v20, s10 :: v_dual_mov_b32 v19, s35
+; GFX11-NEXT:    v_dual_mov_b32 v16, s12 :: v_dual_mov_b32 v21, s34
+; GFX11-NEXT:    v_dual_mov_b32 v12, s14 :: v_dual_mov_b32 v15, s31
+; GFX11-NEXT:    v_dual_mov_b32 v6, s28 :: v_dual_mov_b32 v17, s30
+; GFX11-NEXT:    v_dual_mov_b32 v32, s48 :: v_dual_mov_b32 v11, vcc_hi
+; GFX11-NEXT:    v_dual_mov_b32 v30, s38 :: v_dual_mov_b32 v13, s95
+; GFX11-NEXT:    v_dual_mov_b32 v7, s94 :: v_dual_mov_b32 v2, s92
+; GFX11-NEXT:    v_dual_mov_b32 v9, s93 :: v_dual_mov_b32 v4, s91
+; GFX11-NEXT:  .LBB85_5: ; %end
+; GFX11-NEXT:    s_and_b32 s0, s1, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s89, 8
+; GFX11-NEXT:    s_lshl_b32 s2, s79, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s88, 0xff
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_and_b32 s2, s3, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_lshl_b32 s3, s78, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s77, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s76, 8
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xff, v31
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xff, v18
+; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 8, v25
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v32, 8, v32
+; GFX11-NEXT:    v_or_b32_e32 v28, v31, v28
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xff, v22
+; GFX11-NEXT:    v_or_b32_e32 v18, v18, v25
+; GFX11-NEXT:    v_mov_b32_e32 v25, s1
+; GFX11-NEXT:    v_or_b32_e32 v24, v24, v32
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xff, v29
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
+; GFX11-NEXT:    s_lshl_b32 s1, s75, 8
+; GFX11-NEXT:    s_lshl_b32 s2, s73, 8
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v24
+; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v30
+; GFX11-NEXT:    v_or_b32_e32 v27, v29, v27
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
+; GFX11-NEXT:    v_or_b32_e32 v22, v22, v24
+; GFX11-NEXT:    v_or_b32_e32 v24, v31, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
+; GFX11-NEXT:    s_lshl_b32 s3, s62, 8
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v20
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v16
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX11-NEXT:    v_or_b32_e32 v24, v24, v27
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xff, v23
+; GFX11-NEXT:    v_mov_b32_e32 v23, s0
+; GFX11-NEXT:    s_and_b32 s0, s17, 0xff
+; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 8, v13
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s74, 0xff
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_lshl_b32 s2, s72, 8
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s19, 0xff
+; GFX11-NEXT:    v_lshlrev_b32_e32 v12, 8, v12
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_and_b32 s2, s63, 0xff
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_or_b32_e32 v19, v19, v20
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v17
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_dual_mov_b32 v15, s0 :: v_dual_lshlrev_b32 v20, 16, v15
+; GFX11-NEXT:    v_mov_b32_e32 v17, s1
+; GFX11-NEXT:    s_and_b32 s0, s21, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s61, 8
+; GFX11-NEXT:    s_lshl_b32 s2, s59, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s60, 0xff
+; GFX11-NEXT:    v_or_b32_e32 v8, v8, v13
+; GFX11-NEXT:    v_or_b32_e32 v11, v11, v12
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s23, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s58, 8
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s57, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s56, 8
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v26
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    v_or_b32_e32 v12, v8, v11
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 8, v9
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_or_b32_e32 v26, v27, v26
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v21
+; GFX11-NEXT:    v_mov_b32_e32 v11, s0
+; GFX11-NEXT:    v_mov_b32_e32 v13, s1
+; GFX11-NEXT:    s_and_b32 s0, s25, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s47, 8
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v8
+; GFX11-NEXT:    v_or_b32_e32 v6, v7, v6
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s46, 0xff
+; GFX11-NEXT:    s_lshl_b32 s2, s45, 8
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
+; GFX11-NEXT:    v_or_b32_e32 v21, v14, v21
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    v_or_b32_e32 v4, v1, v4
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v5
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_lshl_b32 s2, s44, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_and_b32 s1, s27, 0xff
+; GFX11-NEXT:    s_and_b32 s3, s43, 0xff
+; GFX11-NEXT:    s_lshl_b32 s4, s42, 8
+; GFX11-NEXT:    v_or_b32_e32 v14, v18, v26
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v21
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    s_or_b32 s2, s3, s4
+; GFX11-NEXT:    v_or_b32_e32 v1, v3, v6
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_or_b32_e32 v16, v18, v19
+; GFX11-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v20
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v2
+; GFX11-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-NEXT:    v_mov_b32_e32 v4, s1
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b128 v0, v[22:25], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[14:17], off offset:16
+; GFX11-NEXT:    scratch_store_b128 v0, v[10:13], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
+; GFX11-NEXT:    v_readlane_b32 s49, v33, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v33, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v33, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v33, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v33, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v33, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v33, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v33, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v33, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v33, 0
+; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT:    scratch_load_b32 v33, off, s32 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -58905,34 +57160,61 @@ define inreg <8 x double> @bitcast_v64i8_to_v8f64_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-LABEL: bitcast_v64i8_to_v8f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v1 :: v_dual_mov_b32 v54, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v14 :: v_dual_mov_b32 v37, v12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v10 :: v_dual_mov_b32 v35, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v6 :: v_dual_mov_b32 v33, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v2 :: v_dual_mov_b32 v31, v0
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, v15 :: v_dual_mov_b32 v32, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v13 :: v_dual_mov_b32 v34, v11
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v12 :: v_dual_mov_b32 v36, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v9 :: v_dual_mov_b32 v38, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v8 :: v_dual_mov_b32 v48, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v5 :: v_dual_mov_b32 v50, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v4 :: v_dual_mov_b32 v52, v2
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v0, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v2, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v65, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)
+; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(12)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(11)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 8, v6
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(10)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 8, v8
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(9)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v10
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(8)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v12
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v14
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(6)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v86
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB87_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
@@ -58969,104 +57251,82 @@ define inreg <8 x double> @bitcast_v64i8_to_v8f64_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-NEXT:    s_and_b32 s8, s9, 0xffff
 ; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s27, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v54
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v31
 ; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v32
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v38
 ; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v83
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v85
 ; GFX11-TRUE16-NEXT:    s_and_b32 s11, s28, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s29, 8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v5
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v37
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v82
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v68
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v69
+; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
 ; GFX11-TRUE16-NEXT:    s_and_b32 s10, s10, 0xffff
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v35
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s10, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v32
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v35
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v33
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v8, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v19
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v37
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v6
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v25
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v13, v14
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v15, v86
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v71
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v80
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v81
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v65
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v67
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v52
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v7
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v64
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v1
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v3
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v30
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v30
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v84
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v82
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v15, v86
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v13, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v68
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v96, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v98, 0xff, v65
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v99, 8, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v14, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v86, v87
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v96, v97
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v98, v99
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v51
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v27
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v29
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v54
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v55
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v21
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v23
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v86, v25
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
@@ -59099,204 +57359,179 @@ define inreg <8 x double> @bitcast_v64i8_to_v8f64_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s16, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s17, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s19, 8
+; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
+; GFX11-TRUE16-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s19, 8
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s21, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s3, s2
+; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s2
+; GFX11-TRUE16-NEXT:    s_or_b32 s2, s4, s3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s3, s22, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s23, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s3
 ; GFX11-TRUE16-NEXT:    s_and_b32 s4, s24, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v54
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s2, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s3, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
 ; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-TRUE16-NEXT:    s_and_b32 s3, s4, 0xffff
+; GFX11-TRUE16-NEXT:    s_and_b32 s4, s26, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s27, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v31
+; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v32
+; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v38
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v49
+; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v22
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v83, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v84, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v85, v2
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v3
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v82, v6
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v39
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v68, v10
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v69, v11
+; GFX11-TRUE16-NEXT:    s_and_b32 s6, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s29, 8
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s6
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v1, v2
+; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v35
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v7, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v34
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, s5, v0
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v34
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v16
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v52
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v70, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v71, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v80, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v81, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v65, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v66, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v67, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v7, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v2, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v17
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v0, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v8, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v21
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v22
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v24
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v8, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v9, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v25
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v1, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v3, v0
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v29
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v1, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v11, v3
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v83
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v64, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v13, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v26
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v28
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v30
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v51
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v50
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v49
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v48
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v39
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v30
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v82
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v12, v0
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v14, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v69
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v12, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v13, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v80
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v70
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v65
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v66
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v64
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v55
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v17, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v19, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v20, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v21, v18
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v27, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v29, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v54, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v55, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v17, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v19, v12
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v21, v14
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v23, v15
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v25, v16
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v11
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v2, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v1, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v14, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v0, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v2, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v17
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v18, v14
 ; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v15, v16
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
@@ -68294,811 +66529,391 @@ define inreg <64 x i8> @bitcast_v32i16_to_v64i8_scalar(<32 x i16> inreg %a, i32
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32i16_to_v64i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s50, 10
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB97_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s94, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s30, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB97_4
-; GFX11-TRUE16-NEXT:  .LBB97_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, s17, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s16, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v22, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v26, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v25, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s19, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s21, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s23, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s25, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s27, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s26, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s24, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s22, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s18, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[17:18]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[28:29], 24, v[21:22]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[9:10]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[23:24], 24, v[13:14]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[29:30], 24, v[25:26]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 24, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 24, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 24, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT:    s_branch .LBB97_5
-; GFX11-TRUE16-NEXT:  .LBB97_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr95_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB97_2
-; GFX11-TRUE16-NEXT:  .LBB97_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v96.l, s50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, s49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v86.l, s48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.l, s39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.l, s38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v83.l, s37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v82.l, s36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v81.l, s35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v80.l, s34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v71.l, s31
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v70.l, s30
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v69.l, s95
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v68.l, s94
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, s93
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, s92
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, s91
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, s90
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, s89
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, s88
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s79
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s78
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s77
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s76
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s75
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s74
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s73
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s72
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s63
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s61
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s59
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s57
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s47
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s4
-; GFX11-TRUE16-NEXT:  .LBB97_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v87, 0xff, v87
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v86
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v85, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v87, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v80, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v81
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v26, v86
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, v85, v84
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v83
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v82, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v22, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v71, v80, v71
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v17, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v69, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v67, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v26, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v21, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v22, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v17, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, v18, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v64
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v55
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v53
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v18, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v23, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v49
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 8, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v23, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v38, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v68, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, v14, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v9, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v10, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v7, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v33
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v18, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v18, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v12, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v7, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v11, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v12, v6
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[25:28], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[66:69], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v40, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v40, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v40, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v40, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v40, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v40, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v40, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v40, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v40, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v32i16_to_v64i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s42, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s49, 9
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB97_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s30, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s31, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB97_4
-; GFX11-FAKE16-NEXT:  .LBB97_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v14, s19, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v13, s18, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v20, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v19, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v24, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v23, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v10, s21, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v6, s23, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v4, s25, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v2, s27, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v1, s26, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v3, s24, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v5, s22, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v9, s20, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[25:26], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[15:16]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[19:20]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[21:22], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[28:29], 24, v[23:24]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 24, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 24, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 8, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 24, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v23
-; GFX11-FAKE16-NEXT:    s_branch .LBB97_5
-; GFX11-FAKE16-NEXT:  .LBB97_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr95
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    s_branch .LBB97_2
-; GFX11-FAKE16-NEXT:  .LBB97_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s0 :: v_dual_mov_b32 v24, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s2 :: v_dual_mov_b32 v20, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s16 :: v_dual_mov_b32 v16, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s18 :: v_dual_mov_b32 v14, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s22 :: v_dual_mov_b32 v6, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s24 :: v_dual_mov_b32 v4, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s26 :: v_dual_mov_b32 v2, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v96, s49 :: v_dual_mov_b32 v87, s39
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s48 :: v_dual_mov_b32 v85, s38
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v84, s37 :: v_dual_mov_b32 v83, s36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, s34 :: v_dual_mov_b32 v81, s35
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v80, s31 :: v_dual_mov_b32 v71, s30
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, vcc_hi :: v_dual_mov_b32 v69, s94
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v68, s95 :: v_dual_mov_b32 v67, s93
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, s92 :: v_dual_mov_b32 v65, s91
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v64, s89 :: v_dual_mov_b32 v55, s90
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, s88 :: v_dual_mov_b32 v53, s79
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s78 :: v_dual_mov_b32 v51, s76
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, s77 :: v_dual_mov_b32 v49, s75
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v48, s74 :: v_dual_mov_b32 v39, s73
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s63 :: v_dual_mov_b32 v37, s72
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s62 :: v_dual_mov_b32 v35, s61
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s60 :: v_dual_mov_b32 v33, s58
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s59 :: v_dual_mov_b32 v31, s57
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s56 :: v_dual_mov_b32 v29, s47
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v7, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s46 :: v_dual_mov_b32 v11, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s44 :: v_dual_mov_b32 v17, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s43 :: v_dual_mov_b32 v21, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, s10 :: v_dual_mov_b32 v26, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s6 :: v_dual_mov_b32 v28, s4
-; GFX11-FAKE16-NEXT:  .LBB97_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v96
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v87, 0xff, v87
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v85, 0xff, v85
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v23, v96
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v87, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v87, 0xff, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v86
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v83
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v71
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, v23, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v27
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v85, v84
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xff, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v27
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v80
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v67, 0xff, v67
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v86
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v19, v83
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v87, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v71
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v70
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v69, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v16, v68
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v67, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v83, v24, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v84, v19, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v85, v20, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v15, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v16, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v65
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v64
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v55
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v54
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v16, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v25, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v51
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v49
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v39
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v19, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v25, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v28, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v13, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v14, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v9, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v10, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v5, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v33
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v9, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v17, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v31
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v22
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v17, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v11, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v12, v8
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[82:85], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[23:26], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v40, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v40, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v40, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v40, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v40, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v40, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v40, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v40, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v32i16_to_v64i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    s_cmp_lg_u32 s28, 0
+; GFX11-NEXT:    s_mov_b32 s42, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v40, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v40, s49, 9
+; GFX11-NEXT:    s_cbranch_scc0 .LBB97_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s44, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s57, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s74, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s75, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s79, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s88, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s89, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s92, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s93, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s95, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s94, s16, 16
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s30, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s31, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s35, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s34, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s36, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s37, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s38, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s49, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-NEXT:    s_cbranch_vccnz .LBB97_4
+; GFX11-NEXT:  .LBB97_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v14, s19, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v13, s18, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v20, s3, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v19, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v24, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v23, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v10, s21, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v6, s23, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v4, s25, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v2, s27, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v1, s26, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v3, s24, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v5, s22, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v9, s20, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_lshrrev_b64 v[25:26], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b64 v[26:27], 24, v[15:16]
+; GFX11-NEXT:    v_lshrrev_b64 v[27:28], 24, v[19:20]
+; GFX11-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[17:18], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b64 v[21:22], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[28:29], 24, v[23:24]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 24, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 24, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 8, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 8, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 24, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 8, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 24, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 8, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 8, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 24, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 8, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 8, v23
+; GFX11-NEXT:    s_branch .LBB97_5
+; GFX11-NEXT:  .LBB97_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr31
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr95
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    s_branch .LBB97_2
+; GFX11-NEXT:  .LBB97_4:
+; GFX11-NEXT:    v_dual_mov_b32 v23, s0 :: v_dual_mov_b32 v24, s1
+; GFX11-NEXT:    v_dual_mov_b32 v19, s2 :: v_dual_mov_b32 v20, s3
+; GFX11-NEXT:    v_dual_mov_b32 v15, s16 :: v_dual_mov_b32 v16, s17
+; GFX11-NEXT:    v_dual_mov_b32 v13, s18 :: v_dual_mov_b32 v14, s19
+; GFX11-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
+; GFX11-NEXT:    v_dual_mov_b32 v5, s22 :: v_dual_mov_b32 v6, s23
+; GFX11-NEXT:    v_dual_mov_b32 v3, s24 :: v_dual_mov_b32 v4, s25
+; GFX11-NEXT:    v_dual_mov_b32 v1, s26 :: v_dual_mov_b32 v2, s27
+; GFX11-NEXT:    v_dual_mov_b32 v96, s49 :: v_dual_mov_b32 v87, s39
+; GFX11-NEXT:    v_dual_mov_b32 v86, s48 :: v_dual_mov_b32 v85, s38
+; GFX11-NEXT:    v_dual_mov_b32 v84, s37 :: v_dual_mov_b32 v83, s36
+; GFX11-NEXT:    v_dual_mov_b32 v82, s34 :: v_dual_mov_b32 v81, s35
+; GFX11-NEXT:    v_dual_mov_b32 v80, s31 :: v_dual_mov_b32 v71, s30
+; GFX11-NEXT:    v_dual_mov_b32 v70, vcc_hi :: v_dual_mov_b32 v69, s94
+; GFX11-NEXT:    v_dual_mov_b32 v68, s95 :: v_dual_mov_b32 v67, s93
+; GFX11-NEXT:    v_dual_mov_b32 v66, s92 :: v_dual_mov_b32 v65, s91
+; GFX11-NEXT:    v_dual_mov_b32 v64, s89 :: v_dual_mov_b32 v55, s90
+; GFX11-NEXT:    v_dual_mov_b32 v54, s88 :: v_dual_mov_b32 v53, s79
+; GFX11-NEXT:    v_dual_mov_b32 v52, s78 :: v_dual_mov_b32 v51, s76
+; GFX11-NEXT:    v_dual_mov_b32 v50, s77 :: v_dual_mov_b32 v49, s75
+; GFX11-NEXT:    v_dual_mov_b32 v48, s74 :: v_dual_mov_b32 v39, s73
+; GFX11-NEXT:    v_dual_mov_b32 v38, s63 :: v_dual_mov_b32 v37, s72
+; GFX11-NEXT:    v_dual_mov_b32 v36, s62 :: v_dual_mov_b32 v35, s61
+; GFX11-NEXT:    v_dual_mov_b32 v34, s60 :: v_dual_mov_b32 v33, s58
+; GFX11-NEXT:    v_dual_mov_b32 v32, s59 :: v_dual_mov_b32 v31, s57
+; GFX11-NEXT:    v_dual_mov_b32 v30, s56 :: v_dual_mov_b32 v29, s47
+; GFX11-NEXT:    v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v7, s40
+; GFX11-NEXT:    v_dual_mov_b32 v18, s46 :: v_dual_mov_b32 v11, s28
+; GFX11-NEXT:    v_dual_mov_b32 v12, s44 :: v_dual_mov_b32 v17, s14
+; GFX11-NEXT:    v_dual_mov_b32 v8, s43 :: v_dual_mov_b32 v21, s12
+; GFX11-NEXT:    v_dual_mov_b32 v25, s10 :: v_dual_mov_b32 v26, s8
+; GFX11-NEXT:    v_dual_mov_b32 v27, s6 :: v_dual_mov_b32 v28, s4
+; GFX11-NEXT:  .LBB97_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xff, v23
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v96, 8, v96
+; GFX11-NEXT:    v_and_b32_e32 v87, 0xff, v87
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
+; GFX11-NEXT:    v_and_b32_e32 v85, 0xff, v85
+; GFX11-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
+; GFX11-NEXT:    v_or_b32_e32 v23, v23, v96
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-NEXT:    v_or_b32_e32 v28, v87, v28
+; GFX11-NEXT:    v_and_b32_e32 v87, 0xff, v82
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshlrev_b32_e32 v86, 8, v86
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v83, 8, v83
+; GFX11-NEXT:    v_lshlrev_b32_e32 v71, 8, v71
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-NEXT:    v_or_b32_e32 v82, v23, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 8, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v81
+; GFX11-NEXT:    v_or_b32_e32 v28, v85, v84
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xff, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v26
+; GFX11-NEXT:    v_or_b32_e32 v20, v20, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v80
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
+; GFX11-NEXT:    v_and_b32_e32 v67, 0xff, v67
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
+; GFX11-NEXT:    v_or_b32_e32 v24, v24, v86
+; GFX11-NEXT:    v_or_b32_e32 v19, v19, v83
+; GFX11-NEXT:    v_or_b32_e32 v23, v87, v23
+; GFX11-NEXT:    v_or_b32_e32 v28, v28, v71
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v70
+; GFX11-NEXT:    v_or_b32_e32 v26, v69, v26
+; GFX11-NEXT:    v_or_b32_e32 v16, v16, v68
+; GFX11-NEXT:    v_or_b32_e32 v66, v67, v66
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
+; GFX11-NEXT:    v_or_b32_e32 v83, v24, v27
+; GFX11-NEXT:    v_or_b32_e32 v84, v19, v23
+; GFX11-NEXT:    v_or_b32_e32 v85, v20, v28
+; GFX11-NEXT:    v_or_b32_e32 v23, v15, v26
+; GFX11-NEXT:    v_or_b32_e32 v24, v16, v66
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v65
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xff, v64
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v25
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v55
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v54
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v53
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v52
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v15
+; GFX11-NEXT:    v_or_b32_e32 v15, v16, v19
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v20
+; GFX11-NEXT:    v_or_b32_e32 v16, v25, v26
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v27
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v51
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v21
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v50
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v49
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v48
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v39
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_or_b32_e32 v19, v19, v20
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v21
+; GFX11-NEXT:    v_or_b32_e32 v20, v25, v26
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v27
+; GFX11-NEXT:    v_or_b32_e32 v17, v28, v17
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-NEXT:    v_or_b32_e32 v25, v13, v15
+; GFX11-NEXT:    v_or_b32_e32 v26, v14, v16
+; GFX11-NEXT:    v_or_b32_e32 v13, v9, v19
+; GFX11-NEXT:    v_or_b32_e32 v14, v10, v20
+; GFX11-NEXT:    v_or_b32_e32 v15, v5, v17
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 8, v37
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 8, v35
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v34
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v33
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v32
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v9, v10
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v16
+; GFX11-NEXT:    v_or_b32_e32 v9, v17, v11
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v31
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v29
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v22
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v19
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v16
+; GFX11-NEXT:    v_or_b32_e32 v7, v17, v7
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v18
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_or_b32_e32 v16, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v1, v3, v9
+; GFX11-NEXT:    v_or_b32_e32 v2, v4, v10
+; GFX11-NEXT:    v_or_b32_e32 v3, v11, v7
+; GFX11-NEXT:    v_or_b32_e32 v4, v12, v8
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b128 v0, v[82:85], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[23:26], off offset:16
+; GFX11-NEXT:    scratch_store_b128 v0, v[13:16], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
+; GFX11-NEXT:    v_readlane_b32 s49, v40, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v40, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v40, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -72881,34 +70696,61 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-LABEL: bitcast_v64i8_to_v32i16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, v1 :: v_dual_mov_b32 v52, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v14 :: v_dual_mov_b32 v31, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v12 :: v_dual_mov_b32 v33, v6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v10 :: v_dual_mov_b32 v35, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v4 :: v_dual_mov_b32 v37, v2
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v15 :: v_dual_mov_b32 v34, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v13 :: v_dual_mov_b32 v54, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v12 :: v_dual_mov_b32 v50, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v11 :: v_dual_mov_b32 v48, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v9 :: v_dual_mov_b32 v36, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v6 :: v_dual_mov_b32 v38, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v39, v2
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v0, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v2, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 8, v0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)
+; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(12)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(11)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v6
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(10)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v8
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(9)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v10
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(8)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v12
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v14
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(6)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB99_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
@@ -72918,138 +70760,170 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
 ; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s17, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v32
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v31
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v53
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v50
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v35
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v3, 0xffff, s9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v39
+; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-TRUE16-NEXT:    v_and_b32_e64 v1, 0xffff, s10
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s27, 8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v38
+; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v31
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v32
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v51
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v5, v50
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v6, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v5, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v49
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v7, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v7, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v24
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v54
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v67
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v8, v23
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v2, 16, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v9, v21
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v3, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v30
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v80
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v82
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v27
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v1, 16, v10
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v66
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v12, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v64
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v55
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v71
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v68
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v83
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v86, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v96, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v65
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v12, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v69
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v83
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v64
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v84
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v25
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v86, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v87, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v96, v96, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v97, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v86, 16, v87
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v96, 16, v97
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v87, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v96, v12, v81
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v97, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v86, v85
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v98, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v0, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v1, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v87
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v96, 16, v97
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v86, 16, v98
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB99_3
 ; GFX11-TRUE16-NEXT:  .LBB99_2: ; %cmp.true
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v68
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v67
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v30
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v65
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v70, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v71, v5
 ; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v66, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v28
 ; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v29, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v27, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v25, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v23, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v34
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v21, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, s24, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s25, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v55, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v32
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v19, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v54, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v17, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
 ; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v53, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v33
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
 ; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
@@ -73058,6 +70932,11 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v31
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v37
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v7
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s7
 ; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
@@ -73075,162 +70954,80 @@ define inreg <32 x i16> @bitcast_v64i8_to_v32i16_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v52, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v38
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v51, v4
 ; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s10
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s9, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v50, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v49, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v35
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v84
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v85
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v82
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v80
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v82
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v64
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v81
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v80
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v71
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v69
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v68
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v66
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v64
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v55
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v29
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v23
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v9, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v49
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v51
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v21, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v22, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v23, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v38
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v21, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v84, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v83, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v48, v8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v39, v4
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v85, v1
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v81, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
 ; GFX11-TRUE16-NEXT:    v_and_b32_e64 v8, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v7, 16, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v21
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v16
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v25, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v16, 16, v17
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v23, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v20, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v21
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v23, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v22, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v19, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v20
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v18, 16, v15
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v2, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v3, 16, v2
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
@@ -76981,704 +74778,358 @@ define inreg <32 x half> @bitcast_v32bf16_to_v32f16_scalar(<32 x bfloat> inreg %
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32bf16_to_v32f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s0
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB103_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB103_4
-; GFX11-TRUE16-NEXT:  .LBB103_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s12, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s12, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s13, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s13, 16
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v4
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s14, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s14, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s27, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v9, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s15, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v6, v2 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s15, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v33
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v3, v4 :: v_dual_add_nc_u32 v3, v6, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v8, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v33, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v9, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v4, v4, v7 :: v_dual_add_nc_u32 v7, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v3, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s16, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v4, v7, v8 :: v_dual_add_nc_u32 v5, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v4, v5, v7 :: v_dual_add_nc_u32 v5, v8, v10
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v10
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v9
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s17, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v11, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v5
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v6, v6, v10 :: v_dual_add_nc_u32 v9, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v9, v10, vcc_lo
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s18, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v10, v6
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s19, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v9, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v12, v10
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v8, v11, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s20, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v8, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v9, v11, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v10
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v13, v8
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s20, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v9, v11 :: v_dual_add_nc_u32 v10, 0x7fff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v9
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s21, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v10, v11, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v12, v13
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s21, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v11, v9
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s22, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v10, v10, v14 :: v_dual_add_nc_u32 v11, 0x7fff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, v15, v12
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v10
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v21, 16, v5
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v11, v14, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v10, v13
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s22, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v15, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v11
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s23, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v10, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v14, v15
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v11, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v15
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, v13, v11
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v26, v14, 16, 1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s23, 16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v11
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v10, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, v26, v14
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s24, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v27, v12, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v14
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, v27, v12
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v27, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s24, 16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v13, v13, v15 :: v_dual_add_nc_u32 v14, 0x7fff, v26
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v26, v27, 16, 1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v13
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s25, 0xffff0000
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v27
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v14, v15, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v26, v27
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v13, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s25, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, v15, v13
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v30, v26, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v27, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s26, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v14, v14, v29 :: v_dual_add_nc_u32 v15, 0x7fff, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v13
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, v30, v26
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v27, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v15, v29, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v26
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s27, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v14, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v15, v15, v29, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s26, 16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v35, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v29, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v29
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v26, v30, 16, 1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v32, vcc_lo
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, v34, v29
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v36, v33
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, v26, v30
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v32, v35
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v30
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v27, v27, v37 :: v_dual_add_nc_u32 v26, 0x7fff, v26
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x7fff, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v34, v38, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v29, v29, v32, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v26, v36, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v30, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v28, 16, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v31, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v32, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v23, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v22, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v18, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v17, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v16, 16, v23
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB103_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB103_2
-; GFX11-TRUE16-NEXT:  .LBB103_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v32bf16_to_v32f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_mov_b32 s15, s3
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, s2
-; GFX11-FAKE16-NEXT:    s_mov_b32 s13, s1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s12, s0
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB103_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB103_4
-; GFX11-FAKE16-NEXT:  .LBB103_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s12, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s12, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s13, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s13, 16
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v7, v4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s14, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s14, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s27, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v9, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s15, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v6, v2 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s15, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v33
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v3, v4 :: v_dual_add_nc_u32 v3, v6, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v8, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v9, v5
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v33, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v4, v7 :: v_dual_add_nc_u32 v7, 0x7fff, v9
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v3, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v4
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s16, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v6
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v4, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v9, v4
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v6
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v5
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v9, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s17, 16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v6, v7 :: v_dual_add_nc_u32 v7, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v9
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v6, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s18, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v11, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v7, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v20
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v7
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v8, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s19, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v9, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v8, v11, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v7
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v12, v10
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v8, 16, 1
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v9, v11, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v10
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v13, v8
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s20, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v9, v11 :: v_dual_add_nc_u32 v10, 0x7fff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v13, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v9
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s21, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v10, v11, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v12, v13
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s21, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, v11, v9
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s22, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v10, v14 :: v_dual_add_nc_u32 v11, 0x7fff, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v9
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v12
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v10
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v11, v14, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v13
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s22, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v15, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v11
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s23, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v14, v15
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v11, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v12
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v15
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v13, v11
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v26, v14, 16, 1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s23, 16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v11
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v12, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, v26, v14
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v12, 16, 1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v14
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, v27, v12
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v27, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s24, 16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v13, v13, v15 :: v_dual_add_nc_u32 v14, 0x7fff, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v26, v27, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v13
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s25, 0xffff0000
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, 0x400000, v27
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v14, v15, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v26, v27
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v13, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s25, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v13
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v30, v26, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v27, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s26, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v29 :: v_dual_add_nc_u32 v15, 0x7fff, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, 0x400000, v13
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, v30, v26
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v27, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v15, v29, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v30
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, 0x400000, v26
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s27, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v14, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v29, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s26, 16
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v35, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v29, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v29
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v26, v30, 16, 1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v32, vcc_lo
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, v34, v29
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v36, v33
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, v26, v30
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, v32, v35
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v30
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v27, v27, v37 :: v_dual_add_nc_u32 v26, 0x7fff, v26
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x7fff, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v35
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v34, v38, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v29, v32, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v26, v36, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v30, 16, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v28, 16, v33
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v31, 16, v30
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v34
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v26, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v32, 16, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v5, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v24, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v23, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v22, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v21, 16, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v18, 16, v19
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v17, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v16, 16, v22
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v23
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB103_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB103_2
-; GFX11-FAKE16-NEXT:  .LBB103_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v32bf16_to_v32f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_mov_b32 s15, s3
+; GFX11-NEXT:    s_mov_b32 s14, s2
+; GFX11-NEXT:    s_mov_b32 s13, s1
+; GFX11-NEXT:    s_mov_b32 s12, s0
+; GFX11-NEXT:    s_cmp_lg_u32 s28, 0
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB103_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB103_4
+; GFX11-NEXT:  .LBB103_2: ; %cmp.true
+; GFX11-NEXT:    s_and_b32 s0, s12, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s1, s12, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s13, 0xffff0000
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s1, s13, 16
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
+; GFX11-NEXT:    v_bfe_u32 v3, v1, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v7, v4, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, v7, v4
+; GFX11-NEXT:    v_bfe_u32 v9, v5, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v1
+; GFX11-NEXT:    s_and_b32 s2, s14, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s0, s14, 16
+; GFX11-NEXT:    s_and_b32 s1, s27, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v3, v8, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v9, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v5
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s15, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v6, v2 :: v_dual_add_nc_u32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_bfe_u32 v6, v7, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s15, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 16, v2
+; GFX11-NEXT:    v_or_b32_e32 v38, 0x400000, v33
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v3, v4 :: v_dual_add_nc_u32 v3, v6, v7
+; GFX11-NEXT:    v_bfe_u32 v4, v8, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v7
+; GFX11-NEXT:    v_bfe_u32 v9, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v8
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, v9, v5
+; GFX11-NEXT:    v_bfe_u32 v36, v33, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v6, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 16, v3
+; GFX11-NEXT:    v_bfe_u32 v3, v6, 16, 1
+; GFX11-NEXT:    s_and_b32 s0, s16, 0xffff0000
+; GFX11-NEXT:    v_dual_cndmask_b32 v4, v4, v7 :: v_dual_add_nc_u32 v7, 0x7fff, v9
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, v3, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v4
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
+; GFX11-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT:    s_lshl_b32 s0, s16, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v9
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v6
+; GFX11-NEXT:    v_bfe_u32 v9, v4, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v5
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    s_and_b32 s0, s17, 0xffff0000
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, v9, v4
+; GFX11-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v6
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v4
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v5
+; GFX11-NEXT:    v_bfe_u32 v10, v9, 16, 1
+; GFX11-NEXT:    s_lshl_b32 s0, s17, 16
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v4, v6, v7 :: v_dual_add_nc_u32 v7, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s18, 0xffff0000
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
+; GFX11-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v9
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, v10, v9
+; GFX11-NEXT:    v_bfe_u32 v11, v6, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 16, v5
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s18, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, v11, v6
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v6
+; GFX11-NEXT:    v_or_b32_e32 v11, 0x400000, v5
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v7, v7, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v10
+; GFX11-NEXT:    v_bfe_u32 v10, v5, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-NEXT:    v_lshl_or_b32 v4, v4, 16, v20
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v7
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s19, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v6, v8, v9, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v10, v5
+; GFX11-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v9, v7, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    s_lshl_b32 s0, s19, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
+; GFX11-NEXT:    v_bfe_u32 v12, v10, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, v9, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_cndmask_b32_e32 v5, v8, v11, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
+; GFX11-NEXT:    v_or_b32_e32 v11, 0x400000, v7
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, v12, v10
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_bfe_u32 v13, v8, 16, 1
+; GFX11-NEXT:    s_and_b32 s0, s20, 0xffff0000
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
+; GFX11-NEXT:    v_cndmask_b32_e32 v7, v9, v11, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v12
+; GFX11-NEXT:    v_or_b32_e32 v11, 0x400000, v10
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, v13, v8
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s20, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v9, v9, v11 :: v_dual_add_nc_u32 v10, 0x7fff, v12
+; GFX11-NEXT:    v_or_b32_e32 v11, 0x400000, v8
+; GFX11-NEXT:    v_bfe_u32 v12, v13, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v9
+; GFX11-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s21, 0xffff0000
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v13
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v10, v11, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, v12, v13
+; GFX11-NEXT:    v_bfe_u32 v11, v9, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    s_lshl_b32 s0, s21, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, v11, v9
+; GFX11-NEXT:    v_bfe_u32 v15, v12, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s22, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v10, v10, v14 :: v_dual_add_nc_u32 v11, 0x7fff, v11
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v9
+; GFX11-NEXT:    v_add_nc_u32_e32 v15, v15, v12
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v10
+; GFX11-NEXT:    v_bfe_u32 v10, v13, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_cndmask_b32_e32 v9, v11, v14, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v15
+; GFX11-NEXT:    v_or_b32_e32 v14, 0x400000, v12
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, v10, v13
+; GFX11-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s22, 16
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v13
+; GFX11-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
+; GFX11-NEXT:    v_bfe_u32 v14, v15, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v11
+; GFX11-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s23, 0xffff0000
+; GFX11-NEXT:    v_cndmask_b32_e32 v10, v10, v12, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v12, v14, v15
+; GFX11-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v13, v11, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
+; GFX11-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v12
+; GFX11-NEXT:    v_or_b32_e32 v12, 0x400000, v15
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, v13, v11
+; GFX11-NEXT:    v_bfe_u32 v26, v14, 16, 1
+; GFX11-NEXT:    s_lshl_b32 s0, s23, 16
+; GFX11-NEXT:    v_or_b32_e32 v15, 0x400000, v11
+; GFX11-NEXT:    v_cndmask_b32_e32 v10, v10, v12, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, v26, v14
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-NEXT:    s_and_b32 s0, s24, 0xffff0000
+; GFX11-NEXT:    v_bfe_u32 v27, v12, 16, 1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v26
+; GFX11-NEXT:    v_or_b32_e32 v15, 0x400000, v14
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, v27, v12
+; GFX11-NEXT:    v_add_f32_e64 v27, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s24, 16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v13, v13, v15 :: v_dual_add_nc_u32 v14, 0x7fff, v26
+; GFX11-NEXT:    v_or_b32_e32 v15, 0x400000, v12
+; GFX11-NEXT:    v_bfe_u32 v26, v27, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v13
+; GFX11-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s25, 0xffff0000
+; GFX11-NEXT:    v_or_b32_e32 v29, 0x400000, v27
+; GFX11-NEXT:    v_cndmask_b32_e32 v12, v14, v15, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v26, v27
+; GFX11-NEXT:    v_bfe_u32 v15, v13, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v26, 0x40c00000, s0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
+; GFX11-NEXT:    s_lshl_b32 s0, s25, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
+; GFX11-NEXT:    v_add_nc_u32_e32 v15, v15, v13
+; GFX11-NEXT:    v_bfe_u32 v30, v26, 16, 1
+; GFX11-NEXT:    v_add_f32_e64 v27, 0x40c00000, s0
+; GFX11-NEXT:    s_and_b32 s0, s26, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v14, v14, v29 :: v_dual_add_nc_u32 v15, 0x7fff, v15
+; GFX11-NEXT:    v_or_b32_e32 v29, 0x400000, v13
+; GFX11-NEXT:    v_add_nc_u32_e32 v30, v30, v26
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
+; GFX11-NEXT:    v_bfe_u32 v14, v27, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v32, 0x400000, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
+; GFX11-NEXT:    v_cndmask_b32_e32 v13, v15, v29, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v30
+; GFX11-NEXT:    v_or_b32_e32 v29, 0x400000, v26
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
+; GFX11-NEXT:    v_add_f32_e64 v30, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s27, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, v14, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
+; GFX11-NEXT:    v_cndmask_b32_e32 v15, v15, v29, vcc_lo
+; GFX11-NEXT:    v_add_f32_e64 v29, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s26, 16
+; GFX11-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
+; GFX11-NEXT:    v_add_f32_e64 v35, 0x40c00000, s0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
+; GFX11-NEXT:    v_bfe_u32 v34, v29, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v37, 0x400000, v29
+; GFX11-NEXT:    v_bfe_u32 v26, v30, 16, 1
+; GFX11-NEXT:    v_cndmask_b32_e32 v14, v14, v32, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, v34, v29
+; GFX11-NEXT:    v_bfe_u32 v32, v35, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, v36, v33
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
+; GFX11-NEXT:    v_add_nc_u32_e32 v26, v26, v30
+; GFX11-NEXT:    v_add_nc_u32_e32 v27, 0x7fff, v27
+; GFX11-NEXT:    v_add_nc_u32_e32 v32, v32, v35
+; GFX11-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
+; GFX11-NEXT:    v_or_b32_e32 v36, 0x400000, v30
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_cndmask_b32 v27, v27, v37 :: v_dual_add_nc_u32 v26, 0x7fff, v26
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
+; GFX11-NEXT:    v_add_nc_u32_e32 v29, 0x7fff, v32
+; GFX11-NEXT:    v_or_b32_e32 v32, 0x400000, v35
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
+; GFX11-NEXT:    v_cndmask_b32_e32 v33, v34, v38, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_cndmask_b32_e32 v29, v29, v32, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v33
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
+; GFX11-NEXT:    v_cndmask_b32_e32 v26, v26, v36, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v15, v30, 16, v27
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v11, v28, 16, v33
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v12, v31, 16, v30
+; GFX11-NEXT:    v_lshl_or_b32 v10, v10, 16, v34
+; GFX11-NEXT:    v_lshl_or_b32 v14, v26, 16, v27
+; GFX11-NEXT:    v_lshl_or_b32 v13, v32, 16, v29
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v6, v5, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v9, v24, 16, v25
+; GFX11-NEXT:    v_lshl_or_b32 v8, v23, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v7, v22, 16, v27
+; GFX11-NEXT:    v_lshl_or_b32 v5, v21, 16, v29
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v3, v18, 16, v19
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v2, v17, 16, v21
+; GFX11-NEXT:    v_lshl_or_b32 v1, v16, 16, v22
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
+; GFX11-NEXT:    v_lshl_or_b32 v0, v0, 16, v23
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB103_3:
+; GFX11-NEXT:    s_branch .LBB103_2
+; GFX11-NEXT:  .LBB103_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
+; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -81402,811 +78853,391 @@ define inreg <64 x i8> @bitcast_v32f16_to_v64i8_scalar(<32 x half> inreg %a, i32
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32f16_to_v64i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s49, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s50, 10
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB105_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s94, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s30, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB105_4
-; GFX11-TRUE16-NEXT:  .LBB105_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v22, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v26, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v25, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s19 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s21 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s23 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s25 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s27 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s26 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s24 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s22 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s18 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[17:18]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[28:29], 24, v[21:22]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[15:16], 24, v[7:8]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[9:10]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[23:24], 24, v[13:14]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[29:30], 24, v[25:26]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 24, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 8, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 24, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 24, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 8, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 24, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT:    s_branch .LBB105_5
-; GFX11-TRUE16-NEXT:  .LBB105_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr95_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB105_2
-; GFX11-TRUE16-NEXT:  .LBB105_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v96.l, s50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, s49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v86.l, s48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.l, s39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.l, s38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v83.l, s37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v82.l, s36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v81.l, s35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v80.l, s34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v71.l, s31
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v70.l, s30
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v69.l, s95
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v68.l, s94
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, s93
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, s92
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, s91
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, s90
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, s89
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, s88
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s79
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s78
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s77
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s76
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s75
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s74
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s73
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s72
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s63
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s62
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s61
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s60
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s59
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s58
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s57
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s56
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s47
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s4
-; GFX11-TRUE16-NEXT:  .LBB105_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xff, v25
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v87, 0xff, v87
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v86
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v96
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v85, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v87, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v80, 0xff, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v25, v29
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v81
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xff, v69
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xff, v67
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v26, v86
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, v85, v84
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v21, v83
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v82, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v22, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v71, v80, v71
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v17, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v69, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v18, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v67, v66
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v81
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v67, 16, v71
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v66
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v26, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v21, v28
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v22, v67
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v66, v17, v68
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, v18, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v64
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v55
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v53
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v18, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v23, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xff, v49
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 8, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v23, v29
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v39
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v38, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v68, v13, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, v14, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v9, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v10, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v7, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v33
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v32
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v9, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v18, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v31
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v19
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v18, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v16
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v12, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v7, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v11, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v12, v6
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[25:28], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[66:69], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v40, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v40, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v40, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v40, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v40, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v40, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v40, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v40, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v40, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v32f16_to_v64i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s42, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s49, 9
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB105_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s94, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s30, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s31, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB105_4
-; GFX11-FAKE16-NEXT:  .LBB105_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v14, 0x200, s19 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v13, 0x200, s18 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v20, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v19, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v24, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v23, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v10, 0x200, s21 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v6, 0x200, s23 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v4, 0x200, s25 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, 0x200, s27 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v1, 0x200, s26 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v3, 0x200, s24 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v5, 0x200, s22 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v9, 0x200, s20 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[25:26], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[15:16]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[19:20]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[21:22], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[28:29], 24, v[23:24]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 24, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 8, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 24, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 8, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 8, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 24, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v23
-; GFX11-FAKE16-NEXT:    s_branch .LBB105_5
-; GFX11-FAKE16-NEXT:  .LBB105_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr95
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    s_branch .LBB105_2
-; GFX11-FAKE16-NEXT:  .LBB105_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s0 :: v_dual_mov_b32 v24, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s2 :: v_dual_mov_b32 v20, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s16 :: v_dual_mov_b32 v16, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s18 :: v_dual_mov_b32 v14, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s22 :: v_dual_mov_b32 v6, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s24 :: v_dual_mov_b32 v4, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s26 :: v_dual_mov_b32 v2, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v96, s49 :: v_dual_mov_b32 v87, s39
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s48 :: v_dual_mov_b32 v85, s38
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v84, s37 :: v_dual_mov_b32 v83, s36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, s34 :: v_dual_mov_b32 v81, s35
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v80, s31 :: v_dual_mov_b32 v71, s30
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, vcc_hi :: v_dual_mov_b32 v69, s94
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v68, s95 :: v_dual_mov_b32 v67, s93
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, s92 :: v_dual_mov_b32 v65, s91
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v64, s89 :: v_dual_mov_b32 v55, s90
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, s88 :: v_dual_mov_b32 v53, s79
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s78 :: v_dual_mov_b32 v51, s76
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, s77 :: v_dual_mov_b32 v49, s75
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v48, s74 :: v_dual_mov_b32 v39, s73
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, s63 :: v_dual_mov_b32 v37, s72
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v36, s62 :: v_dual_mov_b32 v35, s61
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s60 :: v_dual_mov_b32 v33, s58
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s59 :: v_dual_mov_b32 v31, s57
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s56 :: v_dual_mov_b32 v29, s47
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v7, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s46 :: v_dual_mov_b32 v11, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s44 :: v_dual_mov_b32 v17, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s43 :: v_dual_mov_b32 v21, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, s10 :: v_dual_mov_b32 v26, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s6 :: v_dual_mov_b32 v28, s4
-; GFX11-FAKE16-NEXT:  .LBB105_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v96, 8, v96
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v87, 0xff, v87
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v85, 0xff, v85
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v23, v96
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v87, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v87, 0xff, v82
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v86, 8, v86
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v83
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v71
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, v23, v28
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v27
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v81
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v85, v84
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xff, v69
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v20, v27
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v80
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v67, 0xff, v67
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v86
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v19, v83
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v87, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v71
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v70
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v69, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v16, v68
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, v67, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v83, v24, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v84, v19, v23
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v85, v20, v28
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v15, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v16, v66
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v65
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v64
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v55
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v54
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v53
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v52
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v16, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v25, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v51
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 8, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v50
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v49
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v48
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v39
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v38
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v19, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v21
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v25, v26
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v27
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v28, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v13, v15
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v14, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v9, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v10, v20
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v5, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v6
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v37
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v36
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v35
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v33
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v32
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v9, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v3, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v17, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v31
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v22
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v19
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v1, v16
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v17, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v18
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v5, v6
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v9
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v10
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v11, v7
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v12, v8
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[82:85], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[23:26], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v40, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v40, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v40, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v40, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v40, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v40, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v40, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v40, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v32f16_to_v64i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
+; GFX11-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    s_cmp_lg_u32 s28, 0
+; GFX11-NEXT:    s_mov_b32 s42, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v40, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v40, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v40, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v40, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v40, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v40, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v40, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v40, s49, 9
+; GFX11-NEXT:    s_cbranch_scc0 .LBB105_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s44, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s45, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s57, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s58, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s61, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s74, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s75, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s79, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s88, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s89, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s18, 8
+; GFX11-NEXT:    s_lshr_b32 s92, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s93, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s95, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s94, s16, 16
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s30, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s31, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s35, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s34, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s36, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s37, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s38, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s39, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s49, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[28:29], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
+; GFX11-NEXT:    s_cbranch_vccnz .LBB105_4
+; GFX11-NEXT:  .LBB105_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s19 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s18 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s21 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s23 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s25 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s27 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s26 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s24 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s22 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s20 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_lshrrev_b64 v[25:26], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b64 v[26:27], 24, v[15:16]
+; GFX11-NEXT:    v_lshrrev_b64 v[27:28], 24, v[19:20]
+; GFX11-NEXT:    v_lshrrev_b64 v[7:8], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
+; GFX11-NEXT:    v_lshrrev_b64 v[17:18], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b64 v[21:22], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b64 v[28:29], 24, v[23:24]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 24, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 8, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 24, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 8, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 8, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 24, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 8, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 24, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 8, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 8, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 24, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 8, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 8, v23
+; GFX11-NEXT:    s_branch .LBB105_5
+; GFX11-NEXT:  .LBB105_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr49
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr31
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr95
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    s_branch .LBB105_2
+; GFX11-NEXT:  .LBB105_4:
+; GFX11-NEXT:    v_dual_mov_b32 v23, s0 :: v_dual_mov_b32 v24, s1
+; GFX11-NEXT:    v_dual_mov_b32 v19, s2 :: v_dual_mov_b32 v20, s3
+; GFX11-NEXT:    v_dual_mov_b32 v15, s16 :: v_dual_mov_b32 v16, s17
+; GFX11-NEXT:    v_dual_mov_b32 v13, s18 :: v_dual_mov_b32 v14, s19
+; GFX11-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
+; GFX11-NEXT:    v_dual_mov_b32 v5, s22 :: v_dual_mov_b32 v6, s23
+; GFX11-NEXT:    v_dual_mov_b32 v3, s24 :: v_dual_mov_b32 v4, s25
+; GFX11-NEXT:    v_dual_mov_b32 v1, s26 :: v_dual_mov_b32 v2, s27
+; GFX11-NEXT:    v_dual_mov_b32 v96, s49 :: v_dual_mov_b32 v87, s39
+; GFX11-NEXT:    v_dual_mov_b32 v86, s48 :: v_dual_mov_b32 v85, s38
+; GFX11-NEXT:    v_dual_mov_b32 v84, s37 :: v_dual_mov_b32 v83, s36
+; GFX11-NEXT:    v_dual_mov_b32 v82, s34 :: v_dual_mov_b32 v81, s35
+; GFX11-NEXT:    v_dual_mov_b32 v80, s31 :: v_dual_mov_b32 v71, s30
+; GFX11-NEXT:    v_dual_mov_b32 v70, vcc_hi :: v_dual_mov_b32 v69, s94
+; GFX11-NEXT:    v_dual_mov_b32 v68, s95 :: v_dual_mov_b32 v67, s93
+; GFX11-NEXT:    v_dual_mov_b32 v66, s92 :: v_dual_mov_b32 v65, s91
+; GFX11-NEXT:    v_dual_mov_b32 v64, s89 :: v_dual_mov_b32 v55, s90
+; GFX11-NEXT:    v_dual_mov_b32 v54, s88 :: v_dual_mov_b32 v53, s79
+; GFX11-NEXT:    v_dual_mov_b32 v52, s78 :: v_dual_mov_b32 v51, s76
+; GFX11-NEXT:    v_dual_mov_b32 v50, s77 :: v_dual_mov_b32 v49, s75
+; GFX11-NEXT:    v_dual_mov_b32 v48, s74 :: v_dual_mov_b32 v39, s73
+; GFX11-NEXT:    v_dual_mov_b32 v38, s63 :: v_dual_mov_b32 v37, s72
+; GFX11-NEXT:    v_dual_mov_b32 v36, s62 :: v_dual_mov_b32 v35, s61
+; GFX11-NEXT:    v_dual_mov_b32 v34, s60 :: v_dual_mov_b32 v33, s58
+; GFX11-NEXT:    v_dual_mov_b32 v32, s59 :: v_dual_mov_b32 v31, s57
+; GFX11-NEXT:    v_dual_mov_b32 v30, s56 :: v_dual_mov_b32 v29, s47
+; GFX11-NEXT:    v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v7, s40
+; GFX11-NEXT:    v_dual_mov_b32 v18, s46 :: v_dual_mov_b32 v11, s28
+; GFX11-NEXT:    v_dual_mov_b32 v12, s44 :: v_dual_mov_b32 v17, s14
+; GFX11-NEXT:    v_dual_mov_b32 v8, s43 :: v_dual_mov_b32 v21, s12
+; GFX11-NEXT:    v_dual_mov_b32 v25, s10 :: v_dual_mov_b32 v26, s8
+; GFX11-NEXT:    v_dual_mov_b32 v27, s6 :: v_dual_mov_b32 v28, s4
+; GFX11-NEXT:  .LBB105_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xff, v23
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v96, 8, v96
+; GFX11-NEXT:    v_and_b32_e32 v87, 0xff, v87
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 8, v28
+; GFX11-NEXT:    v_and_b32_e32 v85, 0xff, v85
+; GFX11-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
+; GFX11-NEXT:    v_or_b32_e32 v23, v23, v96
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xff, v20
+; GFX11-NEXT:    v_or_b32_e32 v28, v87, v28
+; GFX11-NEXT:    v_and_b32_e32 v87, 0xff, v82
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xff, v24
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshlrev_b32_e32 v86, 8, v86
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v83, 8, v83
+; GFX11-NEXT:    v_lshlrev_b32_e32 v71, 8, v71
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xff, v15
+; GFX11-NEXT:    v_or_b32_e32 v82, v23, v28
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 8, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v81
+; GFX11-NEXT:    v_or_b32_e32 v28, v85, v84
+; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 8, v70
+; GFX11-NEXT:    v_and_b32_e32 v69, 0xff, v69
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v26
+; GFX11-NEXT:    v_or_b32_e32 v20, v20, v27
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v80
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 8, v68
+; GFX11-NEXT:    v_and_b32_e32 v67, 0xff, v67
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 8, v66
+; GFX11-NEXT:    v_or_b32_e32 v24, v24, v86
+; GFX11-NEXT:    v_or_b32_e32 v19, v19, v83
+; GFX11-NEXT:    v_or_b32_e32 v23, v87, v23
+; GFX11-NEXT:    v_or_b32_e32 v28, v28, v71
+; GFX11-NEXT:    v_or_b32_e32 v15, v15, v70
+; GFX11-NEXT:    v_or_b32_e32 v26, v69, v26
+; GFX11-NEXT:    v_or_b32_e32 v16, v16, v68
+; GFX11-NEXT:    v_or_b32_e32 v66, v67, v66
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
+; GFX11-NEXT:    v_or_b32_e32 v83, v24, v27
+; GFX11-NEXT:    v_or_b32_e32 v84, v19, v23
+; GFX11-NEXT:    v_or_b32_e32 v85, v20, v28
+; GFX11-NEXT:    v_or_b32_e32 v23, v15, v26
+; GFX11-NEXT:    v_or_b32_e32 v24, v16, v66
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 8, v65
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xff, v64
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v25
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v55
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v54
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v53
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v52
+; GFX11-NEXT:    v_or_b32_e32 v13, v13, v15
+; GFX11-NEXT:    v_or_b32_e32 v15, v16, v19
+; GFX11-NEXT:    v_or_b32_e32 v14, v14, v20
+; GFX11-NEXT:    v_or_b32_e32 v16, v25, v26
+; GFX11-NEXT:    v_or_b32_e32 v9, v9, v27
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xff, v51
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 8, v21
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v21, 8, v50
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xff, v49
+; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 8, v48
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 8, v39
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xff, v38
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-NEXT:    v_or_b32_e32 v19, v19, v20
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v21
+; GFX11-NEXT:    v_or_b32_e32 v20, v25, v26
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v27
+; GFX11-NEXT:    v_or_b32_e32 v17, v28, v17
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
+; GFX11-NEXT:    v_or_b32_e32 v25, v13, v15
+; GFX11-NEXT:    v_or_b32_e32 v26, v14, v16
+; GFX11-NEXT:    v_or_b32_e32 v13, v9, v19
+; GFX11-NEXT:    v_or_b32_e32 v14, v10, v20
+; GFX11-NEXT:    v_or_b32_e32 v15, v5, v17
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xff, v6
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 8, v37
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xff, v36
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 8, v35
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v34
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v33
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v11
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 8, v32
+; GFX11-NEXT:    v_or_b32_e32 v5, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v6, v9, v10
+; GFX11-NEXT:    v_or_b32_e32 v3, v3, v16
+; GFX11-NEXT:    v_or_b32_e32 v9, v17, v11
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xff, v31
+; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 8, v30
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v16, 8, v29
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xff, v22
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 8, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 8, v18
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xff, v12
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
+; GFX11-NEXT:    v_or_b32_e32 v4, v4, v19
+; GFX11-NEXT:    v_or_b32_e32 v10, v10, v11
+; GFX11-NEXT:    v_or_b32_e32 v1, v1, v16
+; GFX11-NEXT:    v_or_b32_e32 v7, v17, v7
+; GFX11-NEXT:    v_or_b32_e32 v2, v2, v18
+; GFX11-NEXT:    v_or_b32_e32 v8, v12, v8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v2
+; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    v_or_b32_e32 v16, v5, v6
+; GFX11-NEXT:    v_or_b32_e32 v1, v3, v9
+; GFX11-NEXT:    v_or_b32_e32 v2, v4, v10
+; GFX11-NEXT:    v_or_b32_e32 v3, v11, v7
+; GFX11-NEXT:    v_or_b32_e32 v4, v12, v8
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b128 v0, v[82:85], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[23:26], off offset:16
+; GFX11-NEXT:    scratch_store_b128 v0, v[13:16], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:48
+; GFX11-NEXT:    v_readlane_b32 s49, v40, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v40, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v40, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v40, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v40, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v40, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v40, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v40, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
+; GFX11-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -85864,34 +82895,61 @@ define inreg <32 x half> @bitcast_v64i8_to_v32f16_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-LABEL: bitcast_v64i8_to_v32f16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, v1 :: v_dual_mov_b32 v52, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v14 :: v_dual_mov_b32 v31, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v12 :: v_dual_mov_b32 v33, v6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v10 :: v_dual_mov_b32 v35, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v4 :: v_dual_mov_b32 v37, v2
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v15 :: v_dual_mov_b32 v34, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v13 :: v_dual_mov_b32 v54, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v12 :: v_dual_mov_b32 v50, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v11 :: v_dual_mov_b32 v48, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v9 :: v_dual_mov_b32 v36, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v6 :: v_dual_mov_b32 v38, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v39, v2
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v0, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v2, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 8, v0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)
+; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(12)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(11)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v6
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(10)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v8
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(9)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v10
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(8)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v12
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v14
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(6)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB107_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
@@ -85901,138 +82959,170 @@ define inreg <32 x half> @bitcast_v64i8_to_v32f16_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
 ; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s17, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v32
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v31
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v53
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v50
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v35
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v3, 0xffff, s9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v39
+; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-TRUE16-NEXT:    v_and_b32_e64 v1, 0xffff, s10
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s27, 8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v38
+; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v31
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v32
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v51
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v5, v50
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v6, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v5, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v49
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v7, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v7, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v24
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v54
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v67
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v8, v23
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v2, 16, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v9, v21
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v3, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v30
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v80
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v82
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v27
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v1, 16, v10
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v66
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v12, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v64
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v55
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v71
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v68
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v83
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v86, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v96, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v65
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v12, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v69
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v83
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v64
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v84
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v25
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v86, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v87, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v96, v96, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v97, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v86, 16, v87
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v96, 16, v97
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v87, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v96, v12, v81
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v97, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v86, v85
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v98, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v0, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v1, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v87
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v96, 16, v97
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v86, 16, v98
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB107_3
 ; GFX11-TRUE16-NEXT:  .LBB107_2: ; %cmp.true
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v68
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v67
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v30
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v65
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v70, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v71, v5
 ; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v66, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v28
 ; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v29, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v27, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v25, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v23, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v34
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v21, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, s24, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s25, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v55, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v32
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v19, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v54, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v17, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
 ; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v53, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v33
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
 ; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
@@ -86041,6 +83131,11 @@ define inreg <32 x half> @bitcast_v64i8_to_v32f16_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v31
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v37
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v7
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s7
 ; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
@@ -86058,162 +83153,80 @@ define inreg <32 x half> @bitcast_v64i8_to_v32f16_scalar(<64 x i8> inreg %a, i32
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v52, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v38
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v51, v4
 ; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s10
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s9, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v50, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v49, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v35
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v84
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v85
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v82
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v80
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v82
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v64
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v81
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v80
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v71
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v69
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v68
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v65
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v66
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v64
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v55
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v29
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v9, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v49
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v51
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v21, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v22, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v23, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v38
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v21, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v84, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v83, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v48, v8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v39, v4
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v85, v1
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v81, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
 ; GFX11-TRUE16-NEXT:    v_and_b32_e64 v8, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v7, 16, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v21
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v16
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v25, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v16, 16, v17
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v23, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v20, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v21
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v23, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v22, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v19, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v20
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v18, 16, v15
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v2, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v3, 16, v2
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
@@ -91922,1539 +88935,743 @@ define inreg <64 x i8> @bitcast_v32bf16_to_v64i8_scalar(<32 x bfloat> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $sgpr92
 ; GFX9-NEXT:    s_branch .LBB109_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v32bf16_to_v64i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v17, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s30, 0
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 vcc_lo, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s31, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s34, 2
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s35, 3
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s36, 4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s37, 5
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s38, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s39, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s48, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v17, s49, 9
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB109_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s27, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s27, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s26, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s25, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s25, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s24, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s23, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s23, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s22, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s21, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s21, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s20, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s19, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s19, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s94, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s18, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s30, s17, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s17, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s16, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s3, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s3, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], s[26:27], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[24:25], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[22:23], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[20:21], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB109_3
-; GFX11-TRUE16-NEXT:  .LBB109_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s1, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s5, s4, 0x10010
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, s4
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s4, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s4, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s1, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s5, s1, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s5, s1
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s1, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s1, s1, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s43, s1, s46
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s5, s4, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, s4
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s4, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s4, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s5, s0, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, s0
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s0, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s3, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s42, s0, s4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s6, s5, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, s5
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s5, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s5, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s3, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s6, s3, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s6, s3
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s3, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s3, s3, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s2, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s6, s5, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, s5
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s5, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s2, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s6, s2, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, s2
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s2, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s2, s2, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s17, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s6
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s7, s6, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, s6
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s6, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s6, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s8, s7, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s8, s7
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s7, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s6, s7, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s17, s6, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s45, s17, s56
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s8, s7, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s8, s7
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s7, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s6, s7, s6
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s8, s7, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, s7
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s7, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s7, s7, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s19, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s16, s7, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s44, s16, s6
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s9, s8, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s9, s8
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s7, s8, s7
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s7, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s9, s8, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s9, s8
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s7, s8, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s18, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s19, s7, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s31, s19, s57
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s9, s8, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s9, s8
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s7, s8, s7
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s9, s8, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, s8
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s9, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s21, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s18, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s30, s18, s7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s10, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s10, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s10, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s10, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s20, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s21, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s93, s21, s58
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s10, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s10, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s23, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s20, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s92, s20, s10
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s23, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s89, s23, s59
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s25, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s22, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s22, s12
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s25, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s25, s60
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s27, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s24, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s24, s13
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s8, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s11, s9, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s11, s9
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s11, s9, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s26, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s27, s11, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s2, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s3, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 vcc_hi, s27, s61
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s5, s4, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, s4
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s4, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s4, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s6, s5, 0x10010
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s6, s5
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s5, 22
-; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s26, s5, s4
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[76:77], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s26, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[88:89], 24
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 vcc_lo, s26, s49
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[10:11], s[92:93], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[12:13], s[30:31], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[14:15], s[44:45], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[28:29], s[8:9], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[40:41], s[42:43], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, vcc_hi, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, vcc_hi, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s77, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s77, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s76, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s76, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s89, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s89, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s88, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s88, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s93, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s93, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s92, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s92, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s31, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s31, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s94, s30, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s95, s30, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s30, s45, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s45, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s31, s44, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s44, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s9, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s9, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s8, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s43, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s43, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s42, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[8:9], vcc, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, vcc_lo, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, vcc_lo, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s42, 8
-; GFX11-TRUE16-NEXT:  .LBB109_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s49
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s40
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s48
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s46
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s39
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s38
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s28
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s36
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s47
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s35
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s31
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s14
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s45
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s56
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s30
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s95
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s94
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s12
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s93
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s57
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s92
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s9, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s91
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s90
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s10
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s89
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s58
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s88
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s79
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s78
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s59
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s76
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s75
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s4
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s73
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s60
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s72
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s44
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s63
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s61
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s62
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-TRUE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
-; GFX11-TRUE16-NEXT:    s_clause 0x3
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v17, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v17, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v17, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v17, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v17, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v17, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v17, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v17, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v17, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v17, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v17, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB109_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr95_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB109_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v32bf16_to_v64i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v17, s32 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s30, 0
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s28, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 vcc_lo, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s31, 1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s34, 2
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s35, 3
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s36, 4
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s37, 5
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s38, 6
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s39, 7
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s48, 8
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s50, 9
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v17, s51, 10
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB109_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s27, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s27, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s26, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s25, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s25, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s24, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s23, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s23, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s22, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s21, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s21, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s20, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s19, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s19, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s94, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s18, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s17, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s17, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s30, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s31, s16, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s3, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s3, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[26:27], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[24:25], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[22:23], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[20:21], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB109_3
-; GFX11-FAKE16-NEXT:  .LBB109_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s1, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s5, s4, 0x10010
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, s4
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s4, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s4, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s1, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s5, s1, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s5, s1
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s1, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s1, s1, s4
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s43, s1, s46
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s5, s4, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, s4
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s4, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s4, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s5, s0, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, s0
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s0, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s3, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s42, s0, s4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s6, s5, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, s5
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s5, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s6, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s5, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s3, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s6, s3, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s6, s3
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s3, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s3, s3, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s2, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s6, s5, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, s5
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s5, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s6, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s2, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s6, s2, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, s2
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s2, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s6, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s2, s2, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s17, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s7, s6, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, s6
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s6, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s6, s6, s7
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s6, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s8, s7, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s8, s7
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s7, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s6, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s6, s7, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s16, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s17, s6, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s45, s17, s56
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s8, s7, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s8, s7
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s7, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s6, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s6, s7, s6
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s8, s7, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, s7
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s7, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s7, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s19, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s16, s7, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s44, s16, s6
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s9, s8, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s9, s8
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s7, s8, s7
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s7, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s9, s8, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s9, s8
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s7, s8, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, s18, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s19, s7, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s31, s19, s57
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s9, s8, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s9, s8
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s7, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s7, s8, s7
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s9, s8, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, s8
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s8, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s9, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s8, s8, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s21, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s18, s8, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s30, s18, s7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s10, s9, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s10, s9
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s8, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s10, s9, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s10, s9
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s20, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s21, s8, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s93, s21, s58
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s10, s9, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s10, s9
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s8, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s8, s9, s8
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s8, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s10, s9, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, s9
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s9, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s9, s9, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s23, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s20, s9, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s92, s20, s8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s11, s10, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s11, s10
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s10, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s9, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s9, s10, s9
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s9, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s10
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s11, s10, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s11, s10
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s10, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s9, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s9, s10, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s22, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s23, s9, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s89, s23, s59
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s11, s10, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s11, s10
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s10, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s9, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s9, s10, s9
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s9, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s10
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s11, s10, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, s10
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s10, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s11, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s10, s10, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, s25, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s22, s10, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s88, s22, s9
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s12, s11, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s12, s11
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s11, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s10, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s12, s11, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s12, s11
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s11, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, s24, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s25, s10, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s77, s25, s60
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s12, s11, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s12, s11
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s11, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s10, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s12, s11, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s12, s11
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s11, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s11, s27, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s24, s10, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s76, s24, s13
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s12, s11, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s12, s11
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s11, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s10, s11, s10
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s10, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s12, s11, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s12, s11
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s11, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s10, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s12, s11, s10
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, s26, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s27, s12, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s2, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s3, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s51, s27, s61
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s5, s4, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, s4
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s4, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s5, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s4, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s6, s5, 0x10010
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s6, s5
-; GFX11-FAKE16-NEXT:    s_bitset1_b32 s5, 22
-; GFX11-FAKE16-NEXT:    s_addk_i32 s4, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s26, s5, s4
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[76:77], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s26, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[88:89], 24
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s50, s26, s48
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[8:9], s[92:93], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[12:13], s[30:31], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[14:15], s[44:45], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[28:29], s[10:11], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[40:41], s[42:43], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s51, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s51, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s77, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s77, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s76, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s76, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s89, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s89, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s88, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s88, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s93, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s93, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s92, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s92, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s31, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s31, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s94, s30, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s95, s30, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s45, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s45, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s30, s44, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s31, s44, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s11, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s11, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s10, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s43, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s43, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s42, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[10:11], s[50:51], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s50, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s50, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s42, 8
-; GFX11-FAKE16-NEXT:  .LBB109_3: ; %end
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s42, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s48, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s40, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s39, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s46, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s38, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s37, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s36, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s28, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s7, s9
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s35, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s47, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s34, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s3, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s31, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s30, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s14, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s17, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s45, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s56, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, vcc_hi, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s18, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s95, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s94, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s12, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s19, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s93, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s9, s57, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s92, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s7
-; GFX11-FAKE16-NEXT:    s_or_b32 s7, s9, s11
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s91, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s90, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s8, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s21, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s89, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s58, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s88, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s22, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s79, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s78, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s23, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s77, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s7, s59, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s76, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s5, s6
-; GFX11-FAKE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s75, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s74, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s4, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s25, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s73, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s60, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s72, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, s1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s2, s3
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s26, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s44, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s43, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s10, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s27, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s63, 8
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, s61, 0xff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s62, 8
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_or_b32 s5, s6, s7
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s4, 0xffff
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
-; GFX11-FAKE16-NEXT:    s_or_b32 s2, s2, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s4, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
-; GFX11-FAKE16-NEXT:    s_clause 0x3
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:48
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v17, 10
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v17, 9
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v17, 8
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v17, 7
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v17, 6
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v17, 5
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v17, 4
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v17, 3
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v17, 2
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v17, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v17, 0
-; GFX11-FAKE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v17, off, s32 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB109_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr95
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    s_branch .LBB109_2
+; GFX11-LABEL: bitcast_v32bf16_to_v64i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
+; GFX11-NEXT:    scratch_store_b32 off, v17, s32 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s4
+; GFX11-NEXT:    v_writelane_b32 v17, s30, 0
+; GFX11-NEXT:    s_cmp_lg_u32 s28, 0
+; GFX11-NEXT:    s_mov_b32 vcc_lo, 0
+; GFX11-NEXT:    v_writelane_b32 v17, s31, 1
+; GFX11-NEXT:    v_writelane_b32 v17, s34, 2
+; GFX11-NEXT:    v_writelane_b32 v17, s35, 3
+; GFX11-NEXT:    v_writelane_b32 v17, s36, 4
+; GFX11-NEXT:    v_writelane_b32 v17, s37, 5
+; GFX11-NEXT:    v_writelane_b32 v17, s38, 6
+; GFX11-NEXT:    v_writelane_b32 v17, s39, 7
+; GFX11-NEXT:    v_writelane_b32 v17, s48, 8
+; GFX11-NEXT:    v_writelane_b32 v17, s50, 9
+; GFX11-NEXT:    v_writelane_b32 v17, s51, 10
+; GFX11-NEXT:    s_cbranch_scc0 .LBB109_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s62, s27, 24
+; GFX11-NEXT:    s_lshr_b32 s61, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s27, 8
+; GFX11-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s26, 8
+; GFX11-NEXT:    s_lshr_b32 s72, s25, 24
+; GFX11-NEXT:    s_lshr_b32 s60, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s25, 8
+; GFX11-NEXT:    s_lshr_b32 s74, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s24, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s23, 24
+; GFX11-NEXT:    s_lshr_b32 s59, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s23, 8
+; GFX11-NEXT:    s_lshr_b32 s78, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s22, 8
+; GFX11-NEXT:    s_lshr_b32 s88, s21, 24
+; GFX11-NEXT:    s_lshr_b32 s58, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s21, 8
+; GFX11-NEXT:    s_lshr_b32 s90, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s20, 8
+; GFX11-NEXT:    s_lshr_b32 s92, s19, 24
+; GFX11-NEXT:    s_lshr_b32 s57, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s19, 8
+; GFX11-NEXT:    s_lshr_b32 s94, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s95, s18, 8
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s17, 24
+; GFX11-NEXT:    s_lshr_b32 s56, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s17, 8
+; GFX11-NEXT:    s_lshr_b32 s30, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s31, s16, 8
+; GFX11-NEXT:    s_lshr_b32 s34, s3, 24
+; GFX11-NEXT:    s_lshr_b32 s47, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s35, s3, 8
+; GFX11-NEXT:    s_lshr_b32 s36, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s37, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s38, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s39, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s48, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[26:27], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[24:25], 24
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[22:23], 24
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[20:21], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[18:19], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
+; GFX11-NEXT:    s_lshr_b64 s[28:29], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
+; GFX11-NEXT:    s_cbranch_vccnz .LBB109_3
+; GFX11-NEXT:  .LBB109_2: ; %cmp.true
+; GFX11-NEXT:    s_and_b32 s4, s1, 0xffff0000
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s5, s4, 0x10010
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_add_i32 s5, s5, s4
+; GFX11-NEXT:    s_bitset1_b32 s4, 22
+; GFX11-NEXT:    s_addk_i32 s5, 0x7fff
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s4, s4, s5
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s4, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s5, s1, 0x10010
+; GFX11-NEXT:    s_add_i32 s4, s5, s1
+; GFX11-NEXT:    s_bitset1_b32 s1, 22
+; GFX11-NEXT:    s_addk_i32 s4, 0x7fff
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s1, s1, s4
+; GFX11-NEXT:    s_and_b32 s4, s0, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s1, s1, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s4
+; GFX11-NEXT:    s_pack_ll_b32_b16 s43, s1, s46
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s5, s4, 0x10010
+; GFX11-NEXT:    s_add_i32 s5, s5, s4
+; GFX11-NEXT:    s_bitset1_b32 s4, 22
+; GFX11-NEXT:    s_addk_i32 s5, 0x7fff
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s4, s4, s5
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s4, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s5, s0, 0x10010
+; GFX11-NEXT:    s_add_i32 s5, s5, s0
+; GFX11-NEXT:    s_bitset1_b32 s0, 22
+; GFX11-NEXT:    s_addk_i32 s5, 0x7fff
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s0, s0, s5
+; GFX11-NEXT:    s_and_b32 s5, s3, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s42, s0, s4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s6, s5, 0x10010
+; GFX11-NEXT:    s_add_i32 s6, s6, s5
+; GFX11-NEXT:    s_bitset1_b32 s5, 22
+; GFX11-NEXT:    s_addk_i32 s6, 0x7fff
+; GFX11-NEXT:    s_and_b32 s7, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s5, s5, s6
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s5, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s6, s3, 0x10010
+; GFX11-NEXT:    s_add_i32 s5, s6, s3
+; GFX11-NEXT:    s_bitset1_b32 s3, 22
+; GFX11-NEXT:    s_addk_i32 s5, 0x7fff
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s3, s3, s5
+; GFX11-NEXT:    s_and_b32 s5, s2, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s3, s3, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s6, s5, 0x10010
+; GFX11-NEXT:    s_add_i32 s6, s6, s5
+; GFX11-NEXT:    s_bitset1_b32 s5, 22
+; GFX11-NEXT:    s_addk_i32 s6, 0x7fff
+; GFX11-NEXT:    s_and_b32 s7, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s5, s5, s6
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s5, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s2, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s6, s2, 0x10010
+; GFX11-NEXT:    s_add_i32 s6, s6, s2
+; GFX11-NEXT:    s_bitset1_b32 s2, 22
+; GFX11-NEXT:    s_addk_i32 s6, 0x7fff
+; GFX11-NEXT:    s_and_b32 s7, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s2, s2, s6
+; GFX11-NEXT:    s_and_b32 s6, s17, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s2, s2, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s7, s6, 0x10010
+; GFX11-NEXT:    s_add_i32 s7, s7, s6
+; GFX11-NEXT:    s_bitset1_b32 s6, 22
+; GFX11-NEXT:    s_addk_i32 s7, 0x7fff
+; GFX11-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s6, s6, s7
+; GFX11-NEXT:    s_lshl_b32 s7, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s6, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s8, s7, 0x10010
+; GFX11-NEXT:    s_add_i32 s6, s8, s7
+; GFX11-NEXT:    s_bitset1_b32 s7, 22
+; GFX11-NEXT:    s_addk_i32 s6, 0x7fff
+; GFX11-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s6, s7, s6
+; GFX11-NEXT:    s_and_b32 s7, s16, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s17, s6, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s45, s17, s56
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s8, s7, 0x10010
+; GFX11-NEXT:    s_add_i32 s6, s8, s7
+; GFX11-NEXT:    s_bitset1_b32 s7, 22
+; GFX11-NEXT:    s_addk_i32 s6, 0x7fff
+; GFX11-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s6, s7, s6
+; GFX11-NEXT:    s_lshl_b32 s7, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s6, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s8, s7, 0x10010
+; GFX11-NEXT:    s_add_i32 s8, s8, s7
+; GFX11-NEXT:    s_bitset1_b32 s7, 22
+; GFX11-NEXT:    s_addk_i32 s8, 0x7fff
+; GFX11-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s7, s7, s8
+; GFX11-NEXT:    s_and_b32 s8, s19, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s16, s7, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s44, s16, s6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s9, s8, 0x10010
+; GFX11-NEXT:    s_add_i32 s7, s9, s8
+; GFX11-NEXT:    s_bitset1_b32 s8, 22
+; GFX11-NEXT:    s_addk_i32 s7, 0x7fff
+; GFX11-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s7, s8, s7
+; GFX11-NEXT:    s_lshl_b32 s8, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s7, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s9, s8, 0x10010
+; GFX11-NEXT:    s_add_i32 s7, s9, s8
+; GFX11-NEXT:    s_bitset1_b32 s8, 22
+; GFX11-NEXT:    s_addk_i32 s7, 0x7fff
+; GFX11-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s7, s8, s7
+; GFX11-NEXT:    s_and_b32 s8, s18, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s19, s7, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s31, s19, s57
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s9, s8, 0x10010
+; GFX11-NEXT:    s_add_i32 s7, s9, s8
+; GFX11-NEXT:    s_bitset1_b32 s8, 22
+; GFX11-NEXT:    s_addk_i32 s7, 0x7fff
+; GFX11-NEXT:    s_and_b32 s9, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s7, s8, s7
+; GFX11-NEXT:    s_lshl_b32 s8, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s7, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s9, s8, 0x10010
+; GFX11-NEXT:    s_add_i32 s9, s9, s8
+; GFX11-NEXT:    s_bitset1_b32 s8, 22
+; GFX11-NEXT:    s_addk_i32 s9, 0x7fff
+; GFX11-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s8, s8, s9
+; GFX11-NEXT:    s_and_b32 s9, s21, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s18, s8, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s30, s18, s7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s10, s9, 0x10010
+; GFX11-NEXT:    s_add_i32 s8, s10, s9
+; GFX11-NEXT:    s_bitset1_b32 s9, 22
+; GFX11-NEXT:    s_addk_i32 s8, 0x7fff
+; GFX11-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s8, s9, s8
+; GFX11-NEXT:    s_lshl_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s8, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s10, s9, 0x10010
+; GFX11-NEXT:    s_add_i32 s8, s10, s9
+; GFX11-NEXT:    s_bitset1_b32 s9, 22
+; GFX11-NEXT:    s_addk_i32 s8, 0x7fff
+; GFX11-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s8, s9, s8
+; GFX11-NEXT:    s_and_b32 s9, s20, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s21, s8, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s93, s21, s58
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s10, s9, 0x10010
+; GFX11-NEXT:    s_add_i32 s8, s10, s9
+; GFX11-NEXT:    s_bitset1_b32 s9, 22
+; GFX11-NEXT:    s_addk_i32 s8, 0x7fff
+; GFX11-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s8, s9, s8
+; GFX11-NEXT:    s_lshl_b32 s9, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s8, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s9
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s10, s9, 0x10010
+; GFX11-NEXT:    s_add_i32 s10, s10, s9
+; GFX11-NEXT:    s_bitset1_b32 s9, 22
+; GFX11-NEXT:    s_addk_i32 s10, 0x7fff
+; GFX11-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s9, s9, s10
+; GFX11-NEXT:    s_and_b32 s10, s23, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s20, s9, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s92, s20, s8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s11, s10, 0x10010
+; GFX11-NEXT:    s_add_i32 s9, s11, s10
+; GFX11-NEXT:    s_bitset1_b32 s10, 22
+; GFX11-NEXT:    s_addk_i32 s9, 0x7fff
+; GFX11-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s9, s10, s9
+; GFX11-NEXT:    s_lshl_b32 s10, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s9, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s10
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s11, s10, 0x10010
+; GFX11-NEXT:    s_add_i32 s9, s11, s10
+; GFX11-NEXT:    s_bitset1_b32 s10, 22
+; GFX11-NEXT:    s_addk_i32 s9, 0x7fff
+; GFX11-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s9, s10, s9
+; GFX11-NEXT:    s_and_b32 s10, s22, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s23, s9, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s89, s23, s59
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s11, s10, 0x10010
+; GFX11-NEXT:    s_add_i32 s9, s11, s10
+; GFX11-NEXT:    s_bitset1_b32 s10, 22
+; GFX11-NEXT:    s_addk_i32 s9, 0x7fff
+; GFX11-NEXT:    s_and_b32 s11, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s9, s10, s9
+; GFX11-NEXT:    s_lshl_b32 s10, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s9, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s10
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s11, s10, 0x10010
+; GFX11-NEXT:    s_add_i32 s11, s11, s10
+; GFX11-NEXT:    s_bitset1_b32 s10, 22
+; GFX11-NEXT:    s_addk_i32 s11, 0x7fff
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s10, s10, s11
+; GFX11-NEXT:    s_and_b32 s11, s25, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s22, s10, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s88, s22, s9
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s12, s11, 0x10010
+; GFX11-NEXT:    s_add_i32 s10, s12, s11
+; GFX11-NEXT:    s_bitset1_b32 s11, 22
+; GFX11-NEXT:    s_addk_i32 s10, 0x7fff
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s10, s11, s10
+; GFX11-NEXT:    s_lshl_b32 s11, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s10, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s12, s11, 0x10010
+; GFX11-NEXT:    s_add_i32 s10, s12, s11
+; GFX11-NEXT:    s_bitset1_b32 s11, 22
+; GFX11-NEXT:    s_addk_i32 s10, 0x7fff
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s10, s11, s10
+; GFX11-NEXT:    s_and_b32 s11, s24, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s25, s10, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s77, s25, s60
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s12, s11, 0x10010
+; GFX11-NEXT:    s_add_i32 s10, s12, s11
+; GFX11-NEXT:    s_bitset1_b32 s11, 22
+; GFX11-NEXT:    s_addk_i32 s10, 0x7fff
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s10, s11, s10
+; GFX11-NEXT:    s_lshl_b32 s11, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s10, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s12, s11, 0x10010
+; GFX11-NEXT:    s_add_i32 s10, s12, s11
+; GFX11-NEXT:    s_bitset1_b32 s11, 22
+; GFX11-NEXT:    s_addk_i32 s10, 0x7fff
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s10, s11, s10
+; GFX11-NEXT:    s_and_b32 s11, s27, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s24, s10, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s76, s24, s13
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s12, s11, 0x10010
+; GFX11-NEXT:    s_add_i32 s10, s12, s11
+; GFX11-NEXT:    s_bitset1_b32 s11, 22
+; GFX11-NEXT:    s_addk_i32 s10, 0x7fff
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s10, s11, s10
+; GFX11-NEXT:    s_lshl_b32 s11, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s10, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s11
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s12, s11, 0x10010
+; GFX11-NEXT:    s_add_i32 s10, s12, s11
+; GFX11-NEXT:    s_bitset1_b32 s11, 22
+; GFX11-NEXT:    s_addk_i32 s10, 0x7fff
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s12, s11, s10
+; GFX11-NEXT:    s_and_b32 s10, s26, 0xffff0000
+; GFX11-NEXT:    s_lshr_b32 s27, s12, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s2, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s3, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s51, s27, s61
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s5, s4, 0x10010
+; GFX11-NEXT:    s_add_i32 s5, s5, s4
+; GFX11-NEXT:    s_bitset1_b32 s4, 22
+; GFX11-NEXT:    s_addk_i32 s5, 0x7fff
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s4, s4, s5
+; GFX11-NEXT:    s_lshl_b32 s5, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s48, s4, 16
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    s_bfe_u32 s6, s5, 0x10010
+; GFX11-NEXT:    s_add_i32 s4, s6, s5
+; GFX11-NEXT:    s_bitset1_b32 s5, 22
+; GFX11-NEXT:    s_addk_i32 s4, 0x7fff
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s26, s5, s4
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[76:77], 24
+; GFX11-NEXT:    s_lshr_b32 s26, s26, 16
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[88:89], 24
+; GFX11-NEXT:    s_pack_ll_b32_b16 s50, s26, s48
+; GFX11-NEXT:    s_lshr_b64 s[8:9], s[92:93], 24
+; GFX11-NEXT:    s_lshr_b64 s[12:13], s[30:31], 24
+; GFX11-NEXT:    s_lshr_b64 s[14:15], s[44:45], 24
+; GFX11-NEXT:    s_lshr_b64 s[28:29], s[10:11], 24
+; GFX11-NEXT:    s_lshr_b64 s[40:41], s[42:43], 24
+; GFX11-NEXT:    s_lshr_b32 s62, s51, 24
+; GFX11-NEXT:    s_lshr_b32 s63, s51, 8
+; GFX11-NEXT:    s_lshr_b32 s72, s77, 24
+; GFX11-NEXT:    s_lshr_b32 s73, s77, 8
+; GFX11-NEXT:    s_lshr_b32 s74, s76, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s76, 8
+; GFX11-NEXT:    s_lshr_b32 s76, s89, 24
+; GFX11-NEXT:    s_lshr_b32 s77, s89, 8
+; GFX11-NEXT:    s_lshr_b32 s78, s88, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s88, 8
+; GFX11-NEXT:    s_lshr_b32 s88, s93, 24
+; GFX11-NEXT:    s_lshr_b32 s89, s93, 8
+; GFX11-NEXT:    s_lshr_b32 s90, s92, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s92, 8
+; GFX11-NEXT:    s_lshr_b32 s92, s31, 24
+; GFX11-NEXT:    s_lshr_b32 s93, s31, 8
+; GFX11-NEXT:    s_lshr_b32 s94, s30, 16
+; GFX11-NEXT:    s_lshr_b32 s95, s30, 8
+; GFX11-NEXT:    s_lshr_b32 vcc_hi, s45, 24
+; GFX11-NEXT:    s_lshr_b32 s45, s45, 8
+; GFX11-NEXT:    s_lshr_b32 s30, s44, 16
+; GFX11-NEXT:    s_lshr_b32 s31, s44, 8
+; GFX11-NEXT:    s_lshr_b32 s34, s11, 24
+; GFX11-NEXT:    s_lshr_b32 s35, s11, 8
+; GFX11-NEXT:    s_lshr_b32 s36, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s37, s10, 8
+; GFX11-NEXT:    s_lshr_b32 s38, s43, 24
+; GFX11-NEXT:    s_lshr_b32 s39, s43, 8
+; GFX11-NEXT:    s_lshr_b32 s48, s42, 16
+; GFX11-NEXT:    s_lshr_b64 s[10:11], s[50:51], 24
+; GFX11-NEXT:    s_lshr_b32 s43, s50, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s50, 8
+; GFX11-NEXT:    s_lshr_b32 s42, s42, 8
+; GFX11-NEXT:  .LBB109_3: ; %end
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s42, 8
+; GFX11-NEXT:    s_and_b32 s7, s48, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s40, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s5
+; GFX11-NEXT:    s_or_b32 s5, s7, s9
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s39, 8
+; GFX11-NEXT:    s_and_b32 s9, s46, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s38, 8
+; GFX11-NEXT:    s_or_b32 s1, s1, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s5
+; GFX11-NEXT:    s_or_b32 s1, s1, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s37, 8
+; GFX11-NEXT:    s_and_b32 s7, s36, 0xff
+; GFX11-NEXT:    s_lshl_b32 s9, s28, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s5
+; GFX11-NEXT:    s_or_b32 s5, s7, s9
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s35, 8
+; GFX11-NEXT:    s_and_b32 s9, s47, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s34, 8
+; GFX11-NEXT:    s_or_b32 s3, s3, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_and_b32 s3, s3, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s5
+; GFX11-NEXT:    s_or_b32 s3, s3, s7
+; GFX11-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    s_and_b32 s0, s16, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s31, 8
+; GFX11-NEXT:    s_and_b32 s2, s30, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s14, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s17, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s45, 8
+; GFX11-NEXT:    s_and_b32 s5, s56, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, vcc_hi, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s18, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s95, 8
+; GFX11-NEXT:    s_and_b32 s5, s94, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s12, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s5, s19, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s93, 8
+; GFX11-NEXT:    s_and_b32 s9, s57, 0xff
+; GFX11-NEXT:    s_lshl_b32 s11, s92, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s7
+; GFX11-NEXT:    s_or_b32 s7, s9, s11
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s7, s7, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    v_dual_mov_b32 v5, s0 :: v_dual_mov_b32 v6, s1
+; GFX11-NEXT:    v_dual_mov_b32 v7, s2 :: v_dual_mov_b32 v8, s3
+; GFX11-NEXT:    s_and_b32 s0, s20, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s91, 8
+; GFX11-NEXT:    s_and_b32 s2, s90, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s8, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s21, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s89, 8
+; GFX11-NEXT:    s_and_b32 s5, s58, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s88, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s7
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s22, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s79, 8
+; GFX11-NEXT:    s_and_b32 s5, s78, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s6
+; GFX11-NEXT:    s_and_b32 s5, s23, 0xff
+; GFX11-NEXT:    s_lshl_b32 s6, s77, 8
+; GFX11-NEXT:    s_and_b32 s7, s59, 0xff
+; GFX11-NEXT:    s_lshl_b32 s8, s76, 8
+; GFX11-NEXT:    s_or_b32 s5, s5, s6
+; GFX11-NEXT:    s_or_b32 s6, s7, s8
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s5, s5, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s6, s6, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s5, s6
+; GFX11-NEXT:    v_dual_mov_b32 v9, s0 :: v_dual_mov_b32 v10, s1
+; GFX11-NEXT:    v_dual_mov_b32 v11, s2 :: v_dual_mov_b32 v12, s3
+; GFX11-NEXT:    s_and_b32 s0, s24, 0xff
+; GFX11-NEXT:    s_lshl_b32 s1, s75, 8
+; GFX11-NEXT:    s_and_b32 s2, s74, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s4, 8
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s25, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s73, 8
+; GFX11-NEXT:    s_and_b32 s4, s60, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s72, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_or_b32 s0, s0, s1
+; GFX11-NEXT:    s_or_b32 s1, s2, s3
+; GFX11-NEXT:    s_and_b32 s2, s26, 0xff
+; GFX11-NEXT:    s_lshl_b32 s3, s44, 8
+; GFX11-NEXT:    s_and_b32 s4, s43, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s10, 8
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    s_and_b32 s4, s27, 0xff
+; GFX11-NEXT:    s_lshl_b32 s5, s63, 8
+; GFX11-NEXT:    s_and_b32 s6, s61, 0xff
+; GFX11-NEXT:    s_lshl_b32 s7, s62, 8
+; GFX11-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-NEXT:    s_or_b32 s5, s6, s7
+; GFX11-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s3, s3, 16
+; GFX11-NEXT:    s_and_b32 s4, s4, 0xffff
+; GFX11-NEXT:    s_lshl_b32 s5, s5, 16
+; GFX11-NEXT:    s_or_b32 s2, s2, s3
+; GFX11-NEXT:    s_or_b32 s3, s4, s5
+; GFX11-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT:    v_dual_mov_b32 v15, s2 :: v_dual_mov_b32 v16, s3
+; GFX11-NEXT:    s_clause 0x3
+; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off
+; GFX11-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
+; GFX11-NEXT:    scratch_store_b128 v0, v[9:12], off offset:32
+; GFX11-NEXT:    scratch_store_b128 v0, v[13:16], off offset:48
+; GFX11-NEXT:    v_readlane_b32 s51, v17, 10
+; GFX11-NEXT:    v_readlane_b32 s50, v17, 9
+; GFX11-NEXT:    v_readlane_b32 s48, v17, 8
+; GFX11-NEXT:    v_readlane_b32 s39, v17, 7
+; GFX11-NEXT:    v_readlane_b32 s38, v17, 6
+; GFX11-NEXT:    v_readlane_b32 s37, v17, 5
+; GFX11-NEXT:    v_readlane_b32 s36, v17, 4
+; GFX11-NEXT:    v_readlane_b32 s35, v17, 3
+; GFX11-NEXT:    v_readlane_b32 s34, v17, 2
+; GFX11-NEXT:    v_readlane_b32 s31, v17, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v17, 0
+; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT:    scratch_load_b32 v17, off, s32 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB109_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr48
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr39
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr38
+; GFX11-NEXT:    ; implicit-def: $sgpr37
+; GFX11-NEXT:    ; implicit-def: $sgpr36
+; GFX11-NEXT:    ; implicit-def: $sgpr28
+; GFX11-NEXT:    ; implicit-def: $sgpr35
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr34
+; GFX11-NEXT:    ; implicit-def: $sgpr31
+; GFX11-NEXT:    ; implicit-def: $sgpr30
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $vcc_hi
+; GFX11-NEXT:    ; implicit-def: $sgpr95
+; GFX11-NEXT:    ; implicit-def: $sgpr94
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    s_branch .LBB109_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -97204,34 +93421,61 @@ define inreg <32 x bfloat> @bitcast_v64i8_to_v32bf16_scalar(<64 x i8> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v64i8_to_v32bf16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, v1 :: v_dual_mov_b32 v52, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v14 :: v_dual_mov_b32 v31, v8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v12 :: v_dual_mov_b32 v33, v6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v10 :: v_dual_mov_b32 v35, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v4 :: v_dual_mov_b32 v37, v2
 ; GFX11-TRUE16-NEXT:    s_clause 0xf
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:56
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:52
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:48
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:44
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:40
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:36
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:32
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:28
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v0, off, s32 offset:60
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:24
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:20
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:16
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:12
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:8
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:4
-; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v15 :: v_dual_mov_b32 v34, v14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v13 :: v_dual_mov_b32 v54, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v12 :: v_dual_mov_b32 v50, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v11 :: v_dual_mov_b32 v48, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v9 :: v_dual_mov_b32 v36, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v6 :: v_dual_mov_b32 v38, v5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v39, v2
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v0, off, s32 offset:56
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:52
+; GFX11-TRUE16-NEXT:    scratch_load_b32 v2, off, s32 offset:60
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v4, off, s32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v6, off, s32 offset:8
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v8, off, s32 offset:16
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v10, off, s32 offset:24
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v12, off, s32 offset:32
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v14, off, s32 offset:40
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:48
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:44
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:36
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:28
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:20
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:12
+; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:4
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v39, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v49, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v51, 8, v7
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 8, v9
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v53, 8, v11
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v52, 8, v13
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 8, v15
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v55, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 8, v21
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v23
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v25
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v27
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v29
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(15)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v85, 8, v0
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(13)
+; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(12)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 8, v4
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(11)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v6
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(10)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 8, v8
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(9)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v71, 8, v10
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(8)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v83, 8, v12
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v81, 8, v14
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(6)
+; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 8, v84
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB111_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
@@ -97241,138 +93485,170 @@ define inreg <32 x bfloat> @bitcast_v64i8_to_v32bf16_scalar(<64 x i8> inreg %a,
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s3, 8
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s5, s6
 ; GFX11-TRUE16-NEXT:    s_or_b32 s6, s7, s8
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s17, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s16, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s17, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s18, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s19, 8
+; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, s16, 0xff
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s18, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s19, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s6, s6, s7
-; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s20, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s21, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s22, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s23, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v39
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v32
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s8, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s24, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s25, 8
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, s26, 0xff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s27, 8
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, s9
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v31
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v52
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v36
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 8, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v53
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v50
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s7, s8
+; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s22, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s23, 8
+; GFX11-TRUE16-NEXT:    s_or_b32 s7, s7, s8
+; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s24, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s25, 8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v35
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s8
+; GFX11-TRUE16-NEXT:    s_or_b32 s8, s9, s10
 ; GFX11-TRUE16-NEXT:    s_and_b32 s10, s28, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s29, 8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s10, s11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e64 v3, 0xffff, s9
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v39
+; GFX11-TRUE16-NEXT:    s_or_b32 s10, s10, s11
+; GFX11-TRUE16-NEXT:    s_and_b32 s9, s26, 0xff
+; GFX11-TRUE16-NEXT:    v_and_b32_e64 v1, 0xffff, s10
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s27, 8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v38
+; GFX11-TRUE16-NEXT:    s_or_b32 s9, s9, s12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v31
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v33
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v36
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v32
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v51
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v5, v50
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v6, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v53
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v5, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v48
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v54
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v49
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v7, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v34
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v51
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v17
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v7, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v18
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v22
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 8, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v9, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v20
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v21
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v26
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v27
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xff, v24
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xff, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v3, 16, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v34
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v18
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v2
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v54
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v3, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v22
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v67
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v19
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v8, v23
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v28
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v2, 16, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v9, v21
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v3, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v30
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xff, v80
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v82
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v28
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v27
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v1, 16, v10
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v66
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v1, 16, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v12, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v64
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v55
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v71
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v69
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v68
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v65
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v82
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v80
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v12, v13
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v14, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v85
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v83
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v86, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v70
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 8, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v96, 0xff, v84
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 8, v81
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v65
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v3, v66
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v70
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v12, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xff, v69
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v83
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v86, 0xff, v64
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v84
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v25
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v86, v87
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v87, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v96, v96, v97
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v97, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v3, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v86, 16, v87
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v96, 16, v97
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v1, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v87, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v96, v12, v81
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v97, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, v86, v85
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v98, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v0, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v1, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v87
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v96, 16, v97
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v86, 16, v98
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB111_3
 ; GFX11-TRUE16-NEXT:  .LBB111_2: ; %cmp.true
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v68
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v67
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v30
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v65
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v16
 ; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v70, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v71, v5
 ; GFX11-TRUE16-NEXT:    s_and_b32 s4, s28, 0xff
+; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s29, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v66, v6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v28
 ; GFX11-TRUE16-NEXT:    s_or_b32 s4, s5, s4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v29, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v24
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v27, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v25, v6
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v23, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v18
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v34
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v21, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; GFX11-TRUE16-NEXT:    s_and_b32 s5, s24, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s25, 8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v55, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v32
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v19, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v54, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v17, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v9
 ; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v36
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v53, v8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v33
 ; GFX11-TRUE16-NEXT:    s_or_b32 s5, s6, s5
 ; GFX11-TRUE16-NEXT:    s_and_b32 s6, s26, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s27, 8
@@ -97381,6 +93657,11 @@ define inreg <32 x bfloat> @bitcast_v64i8_to_v32bf16_scalar(<64 x i8> inreg %a,
 ; GFX11-TRUE16-NEXT:    s_and_b32 s7, s20, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s21, 8
 ; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v31
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v37
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v7
 ; GFX11-TRUE16-NEXT:    s_or_b32 s7, s8, s7
 ; GFX11-TRUE16-NEXT:    s_and_b32 s8, s22, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s23, 8
@@ -97398,162 +93679,80 @@ define inreg <32 x bfloat> @bitcast_v64i8_to_v32bf16_scalar(<64 x i8> inreg %a,
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 8
 ; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
 ; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s3, 8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v52, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v38
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v51, v4
 ; GFX11-TRUE16-NEXT:    s_or_b32 s10, s11, s10
 ; GFX11-TRUE16-NEXT:    s_or_b32 s0, s1, s0
 ; GFX11-TRUE16-NEXT:    s_or_b32 s1, s3, s2
+; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
+; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s9, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s1, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s10, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v50, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v49, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v20
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v4
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v35
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v84
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s9, s10
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v85
-; GFX11-TRUE16-NEXT:    s_addk_i32 s5, 0x300
-; GFX11-TRUE16-NEXT:    s_addk_i32 s6, 0x300
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v82
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v80
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s7, 0x300
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s8, 0x300
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v82
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v64
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 8, v81
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 8, v83
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v80
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v71
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v70
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 8, v67
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v69
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v68
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v30
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v6, v3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v5, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 8, v65
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v66
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v64
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v22
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v18
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v55
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v26
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v28
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v29
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v24
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 8, v25
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v10
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v9, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v19
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v20
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v34
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v7, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 8, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v33
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v16
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v54
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v7, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff, v16
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v49
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v51
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v7
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v17
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v37
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v53
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v48
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v39
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v50
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 8, v35
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 8, v32
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v36
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v21, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v22, v4
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v23, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v24
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 8, v38
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v52
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v4
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v21, v8
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 8, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s7, s8
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v84, v0
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v83, v2
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v48, v8
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v39, v4
 ; GFX11-TRUE16-NEXT:    s_addk_i32 s4, 0x300
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v85, v1
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v81, v3
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v8, v5
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
 ; GFX11-TRUE16-NEXT:    v_and_b32_e64 v8, 0xffff, s4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v7, 16, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v21
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v16
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v25, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v16, 16, v17
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v23, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v20, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v21
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v23, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v22, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v19, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v20
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v18, 16, v15
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v14, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v2, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v3, 16, v2
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.576bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.576bit.ll
index 3a26a5c263d78..e66762f1e02c2 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.576bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.576bit.ll
@@ -3260,246 +3260,121 @@ define inreg <36 x i16> @bitcast_v18i32_to_v36i16_scalar(<18 x i32> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr10
 ; GFX9-NEXT:    s_branch .LBB13_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v18i32_to_v36i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-TRUE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB13_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s18, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s19, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s18, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s19, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s18, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s19, s4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB13_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB13_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v18i32_to_v36i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-FAKE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB13_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s20, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s24, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s25, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s26, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s27, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s28, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s29, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB13_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    s_branch .LBB13_2
+; GFX11-LABEL: bitcast_v18i32_to_v36i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_mov_b32 s46, 0
+; GFX11-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB13_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s4, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-NEXT:    s_cbranch_vccnz .LBB13_3
+; GFX11-NEXT:  .LBB13_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s4, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 16
+; GFX11-NEXT:  .LBB13_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s20, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s24, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s25, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s26, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s27, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s28, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s29, s4
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
+; GFX11-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
+; GFX11-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
+; GFX11-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
+; GFX11-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
+; GFX11-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB13_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    s_branch .LBB13_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -4730,272 +4605,92 @@ define inreg <18 x i32> @bitcast_v36i16_to_v18i32_scalar(<36 x i16> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_branch .LBB15_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v36i16_to_v18i32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s36, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s37, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s38, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s39, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s48, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s49, 5
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s50, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s51, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s52, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s53, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s54, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s55, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s64, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s65, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s66, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s67, 15
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB15_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s36, s40, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s37, s41, s76
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s48, s56, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s49, s57, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s38, s40, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s39, s41, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s50, s56, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s51, s57, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s40, s40, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s41, s41, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s42, s42, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s43, s43, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s44, s44, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s45, s45, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s46, s46, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s52, s56, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s53, s57, s4
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB15_4
-; GFX11-TRUE16-NEXT:  .LBB15_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s74
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_branch .LBB15_5
-; GFX11-TRUE16-NEXT:  .LBB15_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
-; GFX11-TRUE16-NEXT:    s_branch .LBB15_2
-; GFX11-TRUE16-NEXT:  .LBB15_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
-; GFX11-TRUE16-NEXT:  .LBB15_5: ; %end
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v32, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v32, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v32, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v32, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v32, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v32, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v32, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v32, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v32, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v32, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v32, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v32, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v32, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v32, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v32, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v32, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v36i16_to_v18i32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_mov_b32 s40, 0
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB15_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB15_4
-; GFX11-FAKE16-NEXT:  .LBB15_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB15_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB15_2
-; GFX11-FAKE16-NEXT:  .LBB15_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v36i16_to_v18i32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_mov_b32 s40, 0
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB15_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
+; GFX11-NEXT:    s_cbranch_vccnz .LBB15_4
+; GFX11-NEXT:  .LBB15_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB15_3:
+; GFX11-NEXT:    s_branch .LBB15_2
+; GFX11-NEXT:  .LBB15_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
+; GFX11-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -6335,246 +6030,121 @@ define inreg <36 x half> @bitcast_v18i32_to_v36f16_scalar(<18 x i32> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr10
 ; GFX9-NEXT:    s_branch .LBB17_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v18i32_to_v36f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-TRUE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB17_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s18, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s19, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s18, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s19, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s18, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s19, s4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB17_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB17_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v18i32_to_v36f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-FAKE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB17_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s20, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s24, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s25, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s26, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s27, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s28, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s29, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB17_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    s_branch .LBB17_2
+; GFX11-LABEL: bitcast_v18i32_to_v36f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_mov_b32 s46, 0
+; GFX11-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB17_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s4, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-NEXT:    s_cbranch_vccnz .LBB17_3
+; GFX11-NEXT:  .LBB17_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s4, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 16
+; GFX11-NEXT:  .LBB17_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s20, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s24, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s25, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s26, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s27, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s28, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s29, s4
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
+; GFX11-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
+; GFX11-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
+; GFX11-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
+; GFX11-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
+; GFX11-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB17_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    s_branch .LBB17_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -8005,272 +7575,92 @@ define inreg <18 x i32> @bitcast_v36f16_to_v18i32_scalar(<36 x half> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_branch .LBB19_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v36f16_to_v18i32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s36, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s37, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s38, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s39, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s48, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s49, 5
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s50, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s51, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s52, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s53, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s54, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s55, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s64, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s65, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s66, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s67, 15
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB19_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s36, s40, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s37, s41, s76
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s48, s56, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s49, s57, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s38, s40, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s39, s41, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s50, s56, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s51, s57, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s40, s40, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s41, s41, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s42, s42, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s43, s43, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s44, s44, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s45, s45, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s46, s46, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s52, s56, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s53, s57, s4
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB19_4
-; GFX11-TRUE16-NEXT:  .LBB19_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s74
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_branch .LBB19_5
-; GFX11-TRUE16-NEXT:  .LBB19_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
-; GFX11-TRUE16-NEXT:    s_branch .LBB19_2
-; GFX11-TRUE16-NEXT:  .LBB19_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
-; GFX11-TRUE16-NEXT:  .LBB19_5: ; %end
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v32, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v32, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v32, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v32, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v32, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v32, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v32, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v32, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v32, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v32, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v32, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v32, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v32, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v32, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v32, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v32, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v36f16_to_v18i32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_mov_b32 s40, 0
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB19_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB19_4
-; GFX11-FAKE16-NEXT:  .LBB19_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB19_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB19_2
-; GFX11-FAKE16-NEXT:  .LBB19_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v36f16_to_v18i32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_mov_b32 s40, 0
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB19_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
+; GFX11-NEXT:    s_cbranch_vccnz .LBB19_4
+; GFX11-NEXT:  .LBB19_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB19_3:
+; GFX11-NEXT:    s_branch .LBB19_2
+; GFX11-NEXT:  .LBB19_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
+; GFX11-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -10839,311 +10229,149 @@ define inreg <36 x i16> @bitcast_v18f32_to_v36i16_scalar(<18 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr22
 ; GFX9-NEXT:    s_branch .LBB29_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v18f32_to_v36i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB29_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB29_4
-; GFX11-TRUE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, s29, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, s28, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v15, s27, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, s26, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, s25, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, s24, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, s23, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, s22, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, s21, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, s20, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, s19, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, s18, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, s17, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, s16, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, s3, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, s2, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, s1, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, s0, 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
-; GFX11-TRUE16-NEXT:    s_branch .LBB29_5
-; GFX11-TRUE16-NEXT:  .LBB29_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB29_2
-; GFX11-TRUE16-NEXT:  .LBB29_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s5
-; GFX11-TRUE16-NEXT:  .LBB29_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v35, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v33, 16, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v32, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v31, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v29, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v28, 16, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v27, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v26, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v24, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v23, 16, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v22, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v21, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v19, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v24
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v18f32_to_v36i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB29_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB29_4
-; GFX11-FAKE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, s29, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, s28, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v15, s27, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, s26, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, s25, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, s24, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, s23, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, s22, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, s21, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, s20, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, s19, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, s18, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, s17, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, s16, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, s3, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, s2, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, s1, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, s0, 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
-; GFX11-FAKE16-NEXT:    s_branch .LBB29_5
-; GFX11-FAKE16-NEXT:  .LBB29_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    s_branch .LBB29_2
-; GFX11-FAKE16-NEXT:  .LBB29_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v7, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s18 :: v_dual_mov_b32 v3, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s20 :: v_dual_mov_b32 v11, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v9, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s24 :: v_dual_mov_b32 v17, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s28 :: v_dual_mov_b32 v13, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s14 :: v_dual_mov_b32 v26, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, s12 :: v_dual_mov_b32 v24, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s10 :: v_dual_mov_b32 v22, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v21, s8 :: v_dual_mov_b32 v20, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s6 :: v_dual_mov_b32 v18, s5
-; GFX11-FAKE16-NEXT:  .LBB29_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v35, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v33, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v32, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v31, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v29, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v28, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v27, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v26, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v24, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v23, 16, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v22, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v21, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v19, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v24
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v18f32_to_v36i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB29_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s5, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB29_4
+; GFX11-NEXT:  .LBB29_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f32_e64 v13, s29, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v14, s28, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v15, s27, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v16, s26, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v17, s25, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v8, s24, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v9, s23, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v10, s22, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v11, s21, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v12, s20, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v3, s19, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v4, s18, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v5, s17, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v6, s16, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v7, s3, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v0, s2, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v1, s1, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v2, s0, 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
+; GFX11-NEXT:    s_branch .LBB29_5
+; GFX11-NEXT:  .LBB29_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    s_branch .LBB29_2
+; GFX11-NEXT:  .LBB29_4:
+; GFX11-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v7, s3
+; GFX11-NEXT:    v_dual_mov_b32 v6, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v4, s18 :: v_dual_mov_b32 v3, s19
+; GFX11-NEXT:    v_dual_mov_b32 v12, s20 :: v_dual_mov_b32 v11, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v9, s23
+; GFX11-NEXT:    v_dual_mov_b32 v8, s24 :: v_dual_mov_b32 v17, s25
+; GFX11-NEXT:    v_dual_mov_b32 v16, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v14, s28 :: v_dual_mov_b32 v13, s29
+; GFX11-NEXT:    v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45
+; GFX11-NEXT:    v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43
+; GFX11-NEXT:    v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41
+; GFX11-NEXT:    v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15
+; GFX11-NEXT:    v_dual_mov_b32 v27, s14 :: v_dual_mov_b32 v26, s13
+; GFX11-NEXT:    v_dual_mov_b32 v25, s12 :: v_dual_mov_b32 v24, s11
+; GFX11-NEXT:    v_dual_mov_b32 v23, s10 :: v_dual_mov_b32 v22, s9
+; GFX11-NEXT:    v_dual_mov_b32 v21, s8 :: v_dual_mov_b32 v20, s7
+; GFX11-NEXT:    v_dual_mov_b32 v19, s6 :: v_dual_mov_b32 v18, s5
+; GFX11-NEXT:  .LBB29_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v0, v35, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v2, v33, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v3, v32, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v4, v31, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v6, v29, 16, v33
+; GFX11-NEXT:    v_lshl_or_b32 v7, v28, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshl_or_b32 v8, v27, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v9, v26, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v11, v24, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v12, v23, 16, v29
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
+; GFX11-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v13, v22, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v14, v21, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v19, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v17, v18, 16, v24
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -12374,272 +11602,92 @@ define inreg <18 x float> @bitcast_v36i16_to_v18f32_scalar(<36 x i16> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_branch .LBB31_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v36i16_to_v18f32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s36, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s37, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s38, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s39, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s48, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s49, 5
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s50, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s51, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s52, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s53, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s54, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s55, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s64, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s65, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s66, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s67, 15
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB31_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s36, s40, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s37, s41, s76
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s48, s56, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s49, s57, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s38, s40, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s39, s41, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s50, s56, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s51, s57, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s40, s40, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s41, s41, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s42, s42, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s43, s43, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s44, s44, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s45, s45, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s46, s46, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s52, s56, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s53, s57, s4
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB31_4
-; GFX11-TRUE16-NEXT:  .LBB31_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s74
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_branch .LBB31_5
-; GFX11-TRUE16-NEXT:  .LBB31_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
-; GFX11-TRUE16-NEXT:    s_branch .LBB31_2
-; GFX11-TRUE16-NEXT:  .LBB31_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
-; GFX11-TRUE16-NEXT:  .LBB31_5: ; %end
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v32, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v32, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v32, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v32, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v32, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v32, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v32, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v32, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v32, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v32, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v32, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v32, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v32, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v32, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v32, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v32, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v36i16_to_v18f32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_mov_b32 s40, 0
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB31_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB31_4
-; GFX11-FAKE16-NEXT:  .LBB31_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB31_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB31_2
-; GFX11-FAKE16-NEXT:  .LBB31_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v36i16_to_v18f32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_mov_b32 s40, 0
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB31_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
+; GFX11-NEXT:    s_cbranch_vccnz .LBB31_4
+; GFX11-NEXT:  .LBB31_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB31_3:
+; GFX11-NEXT:    s_branch .LBB31_2
+; GFX11-NEXT:  .LBB31_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
+; GFX11-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -13953,311 +13001,149 @@ define inreg <36 x half> @bitcast_v18f32_to_v36f16_scalar(<18 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr22
 ; GFX9-NEXT:    s_branch .LBB33_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v18f32_to_v36f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB33_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB33_4
-; GFX11-TRUE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, s29, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, s28, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v15, s27, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, s26, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, s25, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, s24, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, s23, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, s22, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, s21, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, s20, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, s19, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, s18, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, s17, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, s16, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, s3, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, s2, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, s1, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, s0, 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
-; GFX11-TRUE16-NEXT:    s_branch .LBB33_5
-; GFX11-TRUE16-NEXT:  .LBB33_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB33_2
-; GFX11-TRUE16-NEXT:  .LBB33_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s5
-; GFX11-TRUE16-NEXT:  .LBB33_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v35, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v33, 16, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v32, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v31, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v29, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v28, 16, v34
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v27, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v26, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v24, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v23, 16, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v22, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v21, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v19, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v24
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v18f32_to_v36f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB33_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB33_4
-; GFX11-FAKE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, s29, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, s28, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v15, s27, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, s26, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, s25, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, s24, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, s23, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, s22, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, s21, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, s20, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, s19, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, s18, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, s17, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, s16, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, s3, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, s2, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, s1, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, s0, 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
-; GFX11-FAKE16-NEXT:    s_branch .LBB33_5
-; GFX11-FAKE16-NEXT:  .LBB33_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    s_branch .LBB33_2
-; GFX11-FAKE16-NEXT:  .LBB33_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v7, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s18 :: v_dual_mov_b32 v3, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s20 :: v_dual_mov_b32 v11, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v9, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s24 :: v_dual_mov_b32 v17, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s28 :: v_dual_mov_b32 v13, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s14 :: v_dual_mov_b32 v26, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, s12 :: v_dual_mov_b32 v24, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s10 :: v_dual_mov_b32 v22, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v21, s8 :: v_dual_mov_b32 v20, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s6 :: v_dual_mov_b32 v18, s5
-; GFX11-FAKE16-NEXT:  .LBB33_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v35, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v33, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v32, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v31, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v29, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v28, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v27, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v26, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v24, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v23, 16, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v22, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v21, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v19, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v24
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v18f32_to_v36f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB33_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s5, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB33_4
+; GFX11-NEXT:  .LBB33_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f32_e64 v13, s29, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v14, s28, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v15, s27, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v16, s26, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v17, s25, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v8, s24, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v9, s23, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v10, s22, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v11, s21, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v12, s20, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v3, s19, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v4, s18, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v5, s17, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v6, s16, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v7, s3, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v0, s2, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v1, s1, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v2, s0, 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
+; GFX11-NEXT:    s_branch .LBB33_5
+; GFX11-NEXT:  .LBB33_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    s_branch .LBB33_2
+; GFX11-NEXT:  .LBB33_4:
+; GFX11-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v7, s3
+; GFX11-NEXT:    v_dual_mov_b32 v6, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v4, s18 :: v_dual_mov_b32 v3, s19
+; GFX11-NEXT:    v_dual_mov_b32 v12, s20 :: v_dual_mov_b32 v11, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v9, s23
+; GFX11-NEXT:    v_dual_mov_b32 v8, s24 :: v_dual_mov_b32 v17, s25
+; GFX11-NEXT:    v_dual_mov_b32 v16, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v14, s28 :: v_dual_mov_b32 v13, s29
+; GFX11-NEXT:    v_dual_mov_b32 v35, s46 :: v_dual_mov_b32 v34, s45
+; GFX11-NEXT:    v_dual_mov_b32 v33, s44 :: v_dual_mov_b32 v32, s43
+; GFX11-NEXT:    v_dual_mov_b32 v31, s42 :: v_dual_mov_b32 v30, s41
+; GFX11-NEXT:    v_dual_mov_b32 v29, s40 :: v_dual_mov_b32 v28, s15
+; GFX11-NEXT:    v_dual_mov_b32 v27, s14 :: v_dual_mov_b32 v26, s13
+; GFX11-NEXT:    v_dual_mov_b32 v25, s12 :: v_dual_mov_b32 v24, s11
+; GFX11-NEXT:    v_dual_mov_b32 v23, s10 :: v_dual_mov_b32 v22, s9
+; GFX11-NEXT:    v_dual_mov_b32 v21, s8 :: v_dual_mov_b32 v20, s7
+; GFX11-NEXT:    v_dual_mov_b32 v19, s6 :: v_dual_mov_b32 v18, s5
+; GFX11-NEXT:  .LBB33_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v0, v35, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v2, v33, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v3, v32, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v4, v31, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v6, v29, 16, v33
+; GFX11-NEXT:    v_lshl_or_b32 v7, v28, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshl_or_b32 v8, v27, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v9, v26, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v11, v24, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v12, v23, 16, v29
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
+; GFX11-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v13, v22, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v14, v21, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v19, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v17, v18, 16, v24
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -15688,272 +14574,92 @@ define inreg <18 x float> @bitcast_v36f16_to_v18f32_scalar(<36 x half> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_branch .LBB35_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v36f16_to_v18f32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s36, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s37, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s38, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s39, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s48, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s49, 5
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s50, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s51, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s52, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s53, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s54, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s55, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s64, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s65, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s66, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s67, 15
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB35_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s36, s40, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s37, s41, s76
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s48, s56, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s49, s57, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s38, s40, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s39, s41, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s50, s56, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s51, s57, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s40, s40, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s41, s41, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s42, s42, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s43, s43, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s44, s44, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s45, s45, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s46, s46, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s52, s56, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s53, s57, s4
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB35_4
-; GFX11-TRUE16-NEXT:  .LBB35_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s74
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_branch .LBB35_5
-; GFX11-TRUE16-NEXT:  .LBB35_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
-; GFX11-TRUE16-NEXT:    s_branch .LBB35_2
-; GFX11-TRUE16-NEXT:  .LBB35_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
-; GFX11-TRUE16-NEXT:  .LBB35_5: ; %end
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v32, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v32, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v32, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v32, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v32, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v32, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v32, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v32, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v32, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v32, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v32, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v32, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v32, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v32, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v32, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v32, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v36f16_to_v18f32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_mov_b32 s40, 0
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB35_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB35_4
-; GFX11-FAKE16-NEXT:  .LBB35_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB35_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB35_2
-; GFX11-FAKE16-NEXT:  .LBB35_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v36f16_to_v18f32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_mov_b32 s40, 0
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB35_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
+; GFX11-NEXT:    s_cbranch_vccnz .LBB35_4
+; GFX11-NEXT:  .LBB35_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB35_3:
+; GFX11-NEXT:    s_branch .LBB35_2
+; GFX11-NEXT:  .LBB35_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
+; GFX11-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -17766,246 +16472,121 @@ define inreg <36 x i16> @bitcast_v9i64_to_v36i16_scalar(<9 x i64> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $sgpr10
 ; GFX9-NEXT:    s_branch .LBB41_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v9i64_to_v36i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-TRUE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB41_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s18, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s19, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s18, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s19, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s18, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s19, s4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB41_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB41_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v9i64_to_v36i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-FAKE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB41_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s20, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s24, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s25, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s26, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s27, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s28, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s29, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB41_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    s_branch .LBB41_2
+; GFX11-LABEL: bitcast_v9i64_to_v36i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_mov_b32 s46, 0
+; GFX11-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB41_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s4, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-NEXT:    s_cbranch_vccnz .LBB41_3
+; GFX11-NEXT:  .LBB41_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s4, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 16
+; GFX11-NEXT:  .LBB41_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s20, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s24, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s25, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s26, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s27, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s28, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s29, s4
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
+; GFX11-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
+; GFX11-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
+; GFX11-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
+; GFX11-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
+; GFX11-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB41_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    s_branch .LBB41_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -19236,272 +17817,92 @@ define inreg <9 x i64> @bitcast_v36i16_to_v9i64_scalar(<36 x i16> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_branch .LBB43_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v36i16_to_v9i64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s36, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s37, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s38, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s39, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s48, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s49, 5
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s50, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s51, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s52, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s53, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s54, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s55, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s64, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s65, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s66, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s67, 15
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB43_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s36, s40, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s37, s41, s76
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s48, s56, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s49, s57, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s38, s40, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s39, s41, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s50, s56, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s51, s57, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s40, s40, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s41, s41, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s42, s42, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s43, s43, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s44, s44, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s45, s45, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s46, s46, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s52, s56, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s53, s57, s4
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB43_4
-; GFX11-TRUE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s74
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_branch .LBB43_5
-; GFX11-TRUE16-NEXT:  .LBB43_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
-; GFX11-TRUE16-NEXT:    s_branch .LBB43_2
-; GFX11-TRUE16-NEXT:  .LBB43_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
-; GFX11-TRUE16-NEXT:  .LBB43_5: ; %end
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v32, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v32, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v32, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v32, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v32, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v32, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v32, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v32, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v32, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v32, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v32, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v32, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v32, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v32, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v32, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v32, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v36i16_to_v9i64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_mov_b32 s40, 0
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB43_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB43_4
-; GFX11-FAKE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB43_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB43_2
-; GFX11-FAKE16-NEXT:  .LBB43_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v36i16_to_v9i64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_mov_b32 s40, 0
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB43_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
+; GFX11-NEXT:    s_cbranch_vccnz .LBB43_4
+; GFX11-NEXT:  .LBB43_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB43_3:
+; GFX11-NEXT:    s_branch .LBB43_2
+; GFX11-NEXT:  .LBB43_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
+; GFX11-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -20851,246 +19252,121 @@ define inreg <36 x half> @bitcast_v9i64_to_v36f16_scalar(<9 x i64> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $sgpr10
 ; GFX9-NEXT:    s_branch .LBB45_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v9i64_to_v36f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-TRUE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB45_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s18, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s19, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s18, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s19, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s18, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s19, s4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB45_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB45_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v9i64_to_v36f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-FAKE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB45_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s20, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s24, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s25, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s26, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s27, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s28, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s29, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB45_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    s_branch .LBB45_2
+; GFX11-LABEL: bitcast_v9i64_to_v36f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_mov_b32 s46, 0
+; GFX11-NEXT:    s_and_b32 s4, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB45_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s4, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-NEXT:    s_cbranch_vccnz .LBB45_3
+; GFX11-NEXT:  .LBB45_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s4, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s0, 16
+; GFX11-NEXT:  .LBB45_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s18, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s19, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s20, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s24, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s25, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s26, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s27, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s28, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s29, s4
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s15 :: v_dual_mov_b32 v7, s14
+; GFX11-NEXT:    v_dual_mov_b32 v8, s13 :: v_dual_mov_b32 v9, s12
+; GFX11-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s10
+; GFX11-NEXT:    v_dual_mov_b32 v12, s9 :: v_dual_mov_b32 v13, s8
+; GFX11-NEXT:    v_dual_mov_b32 v14, s7 :: v_dual_mov_b32 v15, s6
+; GFX11-NEXT:    v_dual_mov_b32 v16, s5 :: v_dual_mov_b32 v17, s4
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB45_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    s_branch .LBB45_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -22521,272 +20797,92 @@ define inreg <9 x i64> @bitcast_v36f16_to_v9i64_scalar(<36 x half> inreg %a, i32
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_branch .LBB47_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v36f16_to_v9i64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s36, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s37, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s38, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s39, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s48, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s49, 5
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s50, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s51, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s52, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s53, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s54, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s55, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s64, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s65, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s66, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s67, 15
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB47_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s36, s40, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s37, s41, s76
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s48, s56, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s49, s57, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s38, s40, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s39, s41, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s50, s56, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s51, s57, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s40, s40, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s41, s41, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s42, s42, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s43, s43, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s44, s44, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s45, s45, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s46, s46, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s52, s56, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s53, s57, s4
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB47_4
-; GFX11-TRUE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s74
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_branch .LBB47_5
-; GFX11-TRUE16-NEXT:  .LBB47_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
-; GFX11-TRUE16-NEXT:    s_branch .LBB47_2
-; GFX11-TRUE16-NEXT:  .LBB47_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
-; GFX11-TRUE16-NEXT:  .LBB47_5: ; %end
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v32, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v32, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v32, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v32, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v32, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v32, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v32, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v32, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v32, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v32, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v32, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v32, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v32, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v32, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v32, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v32, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v36f16_to_v9i64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_mov_b32 s40, 0
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB47_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB47_4
-; GFX11-FAKE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB47_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB47_2
-; GFX11-FAKE16-NEXT:  .LBB47_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v36f16_to_v9i64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_mov_b32 s40, 0
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB47_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
+; GFX11-NEXT:    s_cbranch_vccnz .LBB47_4
+; GFX11-NEXT:  .LBB47_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB47_3:
+; GFX11-NEXT:    s_branch .LBB47_2
+; GFX11-NEXT:  .LBB47_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
+; GFX11-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -23803,293 +21899,140 @@ define inreg <36 x i16> @bitcast_v9f64_to_v36i16_scalar(<9 x double> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $vgpr22
 ; GFX9-NEXT:    s_branch .LBB49_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v9f64_to_v36i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[13:14], s[28:29], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[15:16], s[26:27], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[17:18], s[24:25], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[8:9], s[22:23], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[10:11], s[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[3:4], s[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[5:6], s[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], s[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], s[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v0
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_5
-; GFX11-TRUE16-NEXT:  .LBB49_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s5
-; GFX11-TRUE16-NEXT:  .LBB49_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v34, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v33, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v32, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v27, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v26, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v35, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v31, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v29, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v22, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v24, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v23, 16, v19
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v9f64_to_v36i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[13:14], s[28:29], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[15:16], s[26:27], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[17:18], s[24:25], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[8:9], s[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[10:11], s[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[3:4], s[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[5:6], s[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[19:20], s[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], s[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v0
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_5
-; GFX11-FAKE16-NEXT:  .LBB49_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v19, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v10, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v8, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s24 :: v_dual_mov_b32 v20, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v6, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s28 :: v_dual_mov_b32 v4, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v18, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v16, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v14, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s46 :: v_dual_mov_b32 v7, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s45 :: v_dual_mov_b32 v27, s42
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s44 :: v_dual_mov_b32 v21, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v35, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s40 :: v_dual_mov_b32 v33, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s15 :: v_dual_mov_b32 v31, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s11 :: v_dual_mov_b32 v29, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s8 :: v_dual_mov_b32 v25, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s6 :: v_dual_mov_b32 v23, s5
-; GFX11-FAKE16-NEXT:  .LBB49_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v34, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v33, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v32, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v27, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v26, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v35, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v31, 16, v34
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v29, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v25, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v22, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v24, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v23, 16, v19
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v9f64_to_v36i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s5, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_4
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[13:14], s[28:29], 1.0
+; GFX11-NEXT:    v_add_f64 v[15:16], s[26:27], 1.0
+; GFX11-NEXT:    v_add_f64 v[17:18], s[24:25], 1.0
+; GFX11-NEXT:    v_add_f64 v[8:9], s[22:23], 1.0
+; GFX11-NEXT:    v_add_f64 v[10:11], s[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[3:4], s[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[5:6], s[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[19:20], s[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], s[0:1], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v0
+; GFX11-NEXT:    s_branch .LBB49_5
+; GFX11-NEXT:  .LBB49_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    s_branch .LBB49_2
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v19, s2
+; GFX11-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v10, s20
+; GFX11-NEXT:    v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v8, s22
+; GFX11-NEXT:    v_dual_mov_b32 v17, s24 :: v_dual_mov_b32 v20, s3
+; GFX11-NEXT:    v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v6, s17
+; GFX11-NEXT:    v_dual_mov_b32 v13, s28 :: v_dual_mov_b32 v4, s19
+; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v18, s25
+; GFX11-NEXT:    v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v16, s27
+; GFX11-NEXT:    v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v14, s29
+; GFX11-NEXT:    v_dual_mov_b32 v34, s46 :: v_dual_mov_b32 v7, s43
+; GFX11-NEXT:    v_dual_mov_b32 v2, s45 :: v_dual_mov_b32 v27, s42
+; GFX11-NEXT:    v_dual_mov_b32 v30, s44 :: v_dual_mov_b32 v21, s14
+; GFX11-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v35, s13
+; GFX11-NEXT:    v_dual_mov_b32 v12, s40 :: v_dual_mov_b32 v33, s12
+; GFX11-NEXT:    v_dual_mov_b32 v22, s15 :: v_dual_mov_b32 v31, s10
+; GFX11-NEXT:    v_dual_mov_b32 v32, s11 :: v_dual_mov_b32 v29, s9
+; GFX11-NEXT:    v_dual_mov_b32 v28, s8 :: v_dual_mov_b32 v25, s7
+; GFX11-NEXT:    v_dual_mov_b32 v24, s6 :: v_dual_mov_b32 v23, s5
+; GFX11-NEXT:  .LBB49_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshl_or_b32 v2, v2, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v0, v34, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v3, v33, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v4
+; GFX11-NEXT:    v_lshl_or_b32 v4, v30, 16, v5
+; GFX11-NEXT:    v_lshl_or_b32 v5, v32, 16, v6
+; GFX11-NEXT:    v_lshl_or_b32 v6, v7, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v9
+; GFX11-NEXT:    v_lshl_or_b32 v8, v27, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v26, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v12, v12, 16, v17
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v1, v35, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v7, v31, 16, v34
+; GFX11-NEXT:    v_lshl_or_b32 v9, v29, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v11, v28, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v13, v25, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v14, v22, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v15, v24, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v16, v21, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v17, v23, 16, v19
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -25320,272 +23263,92 @@ define inreg <9 x double> @bitcast_v36i16_to_v9f64_scalar(<36 x i16> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_branch .LBB51_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v36i16_to_v9f64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s36, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s37, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s38, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s39, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s48, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s49, 5
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s50, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s51, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s52, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s53, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s54, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s55, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s64, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s65, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s66, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s67, 15
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB51_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s36, s40, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s37, s41, s76
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s48, s56, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s49, s57, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s38, s40, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s39, s41, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s50, s56, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s51, s57, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s40, s40, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s41, s41, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s42, s42, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s43, s43, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s44, s44, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s45, s45, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s46, s46, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s52, s56, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s53, s57, s4
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB51_4
-; GFX11-TRUE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s74
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_branch .LBB51_5
-; GFX11-TRUE16-NEXT:  .LBB51_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
-; GFX11-TRUE16-NEXT:    s_branch .LBB51_2
-; GFX11-TRUE16-NEXT:  .LBB51_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
-; GFX11-TRUE16-NEXT:  .LBB51_5: ; %end
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v32, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v32, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v32, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v32, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v32, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v32, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v32, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v32, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v32, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v32, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v32, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v32, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v32, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v32, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v32, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v32, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v36i16_to_v9f64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_mov_b32 s40, 0
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB51_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB51_4
-; GFX11-FAKE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB51_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB51_2
-; GFX11-FAKE16-NEXT:  .LBB51_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v36i16_to_v9f64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_mov_b32 s40, 0
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB51_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
+; GFX11-NEXT:    s_cbranch_vccnz .LBB51_4
+; GFX11-NEXT:  .LBB51_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB51_3:
+; GFX11-NEXT:    s_branch .LBB51_2
+; GFX11-NEXT:  .LBB51_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
+; GFX11-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -26836,293 +24599,140 @@ define inreg <36 x half> @bitcast_v9f64_to_v36f16_scalar(<9 x double> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr22
 ; GFX9-NEXT:    s_branch .LBB53_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v9f64_to_v36f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB53_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB53_4
-; GFX11-TRUE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[13:14], s[28:29], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[15:16], s[26:27], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[17:18], s[24:25], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[8:9], s[22:23], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[10:11], s[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[3:4], s[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[5:6], s[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], s[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], s[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v0
-; GFX11-TRUE16-NEXT:    s_branch .LBB53_5
-; GFX11-TRUE16-NEXT:  .LBB53_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB53_2
-; GFX11-TRUE16-NEXT:  .LBB53_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s46
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s5
-; GFX11-TRUE16-NEXT:  .LBB53_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v34, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v33, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v32, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v27, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v26, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v35, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v31, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v29, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v22, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v24, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v23, 16, v19
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v9f64_to_v36f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB53_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB53_4
-; GFX11-FAKE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[13:14], s[28:29], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[15:16], s[26:27], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[17:18], s[24:25], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[8:9], s[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[10:11], s[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[3:4], s[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[5:6], s[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[19:20], s[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], s[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v0
-; GFX11-FAKE16-NEXT:    s_branch .LBB53_5
-; GFX11-FAKE16-NEXT:  .LBB53_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    s_branch .LBB53_2
-; GFX11-FAKE16-NEXT:  .LBB53_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v19, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v10, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v8, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s24 :: v_dual_mov_b32 v20, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v6, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s28 :: v_dual_mov_b32 v4, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v18, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v16, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v14, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s46 :: v_dual_mov_b32 v7, s43
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s45 :: v_dual_mov_b32 v27, s42
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s44 :: v_dual_mov_b32 v21, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v35, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s40 :: v_dual_mov_b32 v33, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s15 :: v_dual_mov_b32 v31, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s11 :: v_dual_mov_b32 v29, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s8 :: v_dual_mov_b32 v25, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s6 :: v_dual_mov_b32 v23, s5
-; GFX11-FAKE16-NEXT:  .LBB53_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v34, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v33, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v32, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v27, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v26, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v35, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v31, 16, v34
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v29, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v25, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v22, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v24, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v23, 16, v19
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v9f64_to_v36f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_and_b32 s5, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB53_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s5, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB53_4
+; GFX11-NEXT:  .LBB53_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[13:14], s[28:29], 1.0
+; GFX11-NEXT:    v_add_f64 v[15:16], s[26:27], 1.0
+; GFX11-NEXT:    v_add_f64 v[17:18], s[24:25], 1.0
+; GFX11-NEXT:    v_add_f64 v[8:9], s[22:23], 1.0
+; GFX11-NEXT:    v_add_f64 v[10:11], s[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[3:4], s[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[5:6], s[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[19:20], s[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], s[0:1], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v0
+; GFX11-NEXT:    s_branch .LBB53_5
+; GFX11-NEXT:  .LBB53_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    s_branch .LBB53_2
+; GFX11-NEXT:  .LBB53_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v19, s2
+; GFX11-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v10, s20
+; GFX11-NEXT:    v_dual_mov_b32 v3, s18 :: v_dual_mov_b32 v8, s22
+; GFX11-NEXT:    v_dual_mov_b32 v17, s24 :: v_dual_mov_b32 v20, s3
+; GFX11-NEXT:    v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v6, s17
+; GFX11-NEXT:    v_dual_mov_b32 v13, s28 :: v_dual_mov_b32 v4, s19
+; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v18, s25
+; GFX11-NEXT:    v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v16, s27
+; GFX11-NEXT:    v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v14, s29
+; GFX11-NEXT:    v_dual_mov_b32 v34, s46 :: v_dual_mov_b32 v7, s43
+; GFX11-NEXT:    v_dual_mov_b32 v2, s45 :: v_dual_mov_b32 v27, s42
+; GFX11-NEXT:    v_dual_mov_b32 v30, s44 :: v_dual_mov_b32 v21, s14
+; GFX11-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v35, s13
+; GFX11-NEXT:    v_dual_mov_b32 v12, s40 :: v_dual_mov_b32 v33, s12
+; GFX11-NEXT:    v_dual_mov_b32 v22, s15 :: v_dual_mov_b32 v31, s10
+; GFX11-NEXT:    v_dual_mov_b32 v32, s11 :: v_dual_mov_b32 v29, s9
+; GFX11-NEXT:    v_dual_mov_b32 v28, s8 :: v_dual_mov_b32 v25, s7
+; GFX11-NEXT:    v_dual_mov_b32 v24, s6 :: v_dual_mov_b32 v23, s5
+; GFX11-NEXT:  .LBB53_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshl_or_b32 v2, v2, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v0, v34, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v3, v33, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v4
+; GFX11-NEXT:    v_lshl_or_b32 v4, v30, 16, v5
+; GFX11-NEXT:    v_lshl_or_b32 v5, v32, 16, v6
+; GFX11-NEXT:    v_lshl_or_b32 v6, v7, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v9
+; GFX11-NEXT:    v_lshl_or_b32 v8, v27, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v26, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v12, v12, 16, v17
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v1, v35, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v7, v31, 16, v34
+; GFX11-NEXT:    v_lshl_or_b32 v9, v29, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v11, v28, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v13, v25, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v14, v22, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v15, v24, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v16, v21, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v17, v23, 16, v19
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -28553,272 +26163,92 @@ define inreg <9 x double> @bitcast_v36f16_to_v9f64_scalar(<36 x half> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
 ; GFX9-NEXT:    s_branch .LBB55_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v36f16_to_v9f64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s4, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s36, 0
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s37, 1
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s38, 2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s39, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s48, 4
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s49, 5
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s50, 6
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s51, 7
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s52, 8
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s53, 9
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s54, 10
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s55, 11
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s64, 12
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s65, 13
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s66, 14
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v32, s67, 15
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB55_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s36, s40, s77
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s37, s41, s76
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s48, s56, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s49, s57, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s38, s40, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s39, s41, s74
-; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s50, s56, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s51, s57, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s40, s40, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s41, s41, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s42, s42, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s43, s43, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s44, s44, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s45, s45, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s46, s46, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s52, s56, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s53, s57, s4
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB55_4
-; GFX11-TRUE16-NEXT:  .LBB55_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s74
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_branch .LBB55_5
-; GFX11-TRUE16-NEXT:  .LBB55_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
-; GFX11-TRUE16-NEXT:    s_branch .LBB55_2
-; GFX11-TRUE16-NEXT:  .LBB55_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
-; GFX11-TRUE16-NEXT:  .LBB55_5: ; %end
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v32, 15
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v32, 14
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v32, 13
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v32, 12
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v32, 11
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v32, 10
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v32, 9
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v32, 8
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v32, 7
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v32, 6
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v32, 5
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v32, 4
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v32, 3
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v32, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v32, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v32, 0
-; GFX11-TRUE16-NEXT:    s_xor_saveexec_b32 s0, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v36f16_to_v9f64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s0, 16
-; GFX11-FAKE16-NEXT:    s_mov_b32 s40, 0
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB55_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB55_4
-; GFX11-FAKE16-NEXT:  .LBB55_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB55_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB55_2
-; GFX11-FAKE16-NEXT:  .LBB55_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v36f16_to_v9f64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-NEXT:    s_mov_b32 s40, 0
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB55_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
+; GFX11-NEXT:    s_cbranch_vccnz .LBB55_4
+; GFX11-NEXT:  .LBB55_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB55_3:
+; GFX11-NEXT:    s_branch .LBB55_2
+; GFX11-NEXT:  .LBB55_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s15
+; GFX11-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT:    v_dual_mov_b32 v22, s22 :: v_dual_mov_b32 v23, s23
+; GFX11-NEXT:    v_dual_mov_b32 v24, s24 :: v_dual_mov_b32 v25, s25
+; GFX11-NEXT:    v_dual_mov_b32 v26, s26 :: v_dual_mov_b32 v27, s27
+; GFX11-NEXT:    v_dual_mov_b32 v28, s28 :: v_dual_mov_b32 v29, s29
+; GFX11-NEXT:    v_dual_mov_b32 v30, s30 :: v_dual_mov_b32 v31, s31
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -30228,317 +27658,149 @@ define inreg <36 x half> @bitcast_v36i16_to_v36f16_scalar(<36 x i16> inreg %a, i
 ; GFX9-NEXT:    v_mov_b32_e32 v3, v19
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v36i16_to_v36f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB57_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB57_4
-; GFX11-TRUE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s15, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s15, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s14, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s13, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s12, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s11, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s10, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s29, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s28, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s26, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s25, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s24, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v13
-; GFX11-TRUE16-NEXT:    s_branch .LBB57_5
-; GFX11-TRUE16-NEXT:  .LBB57_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB57_2
-; GFX11-TRUE16-NEXT:  .LBB57_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s5
-; GFX11-TRUE16-NEXT:  .LBB57_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v28, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v29, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v23, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v24, 16, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v27, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v18, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v22, 16, v24
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v36i16_to_v36f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s0, 16
-; GFX11-FAKE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB57_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB57_4
-; GFX11-FAKE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s21, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s20, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s19, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s18, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s17, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v13, s29, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v14, s28, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v16, s26, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v17, s25, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v8, s24, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v9, s15, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v10, s14, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v11, s13, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v3, s11, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v4, s10, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v5, s9, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v0, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v7, s3, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
-; GFX11-FAKE16-NEXT:    s_branch .LBB57_5
-; GFX11-FAKE16-NEXT:  .LBB57_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB57_2
-; GFX11-FAKE16-NEXT:  .LBB57_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s29 :: v_dual_mov_b32 v14, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v16, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s25 :: v_dual_mov_b32 v8, s24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v10, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v12, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s19 :: v_dual_mov_b32 v4, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s3 :: v_dual_mov_b32 v0, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s45 :: v_dual_mov_b32 v19, s44
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s43 :: v_dual_mov_b32 v21, s42
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s41 :: v_dual_mov_b32 v23, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s15 :: v_dual_mov_b32 v25, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s11 :: v_dual_mov_b32 v29, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s9 :: v_dual_mov_b32 v31, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s6 :: v_dual_mov_b32 v33, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s4 :: v_dual_mov_b32 v35, s5
-; GFX11-FAKE16-NEXT:  .LBB57_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v35, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v33, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v32, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v31, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v29, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v28, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v27, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v26, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v24, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v23, 16, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v22, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v21, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v19, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v24
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v36i16_to_v36f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_lshr_b32 s45, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s0, 16
+; GFX11-NEXT:    s_mov_b32 s46, 0
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB57_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-NEXT:    s_cbranch_vccnz .LBB57_4
+; GFX11-NEXT:  .LBB57_2: ; %cmp.true
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s21, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s20, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s19, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s18, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s17, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-NEXT:    v_pk_add_u16 v13, s29, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v14, s28, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v16, s26, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v17, s25, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v8, s24, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v9, s15, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v10, s14, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v11, s13, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v3, s11, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v4, s10, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v5, s9, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v0, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v7, s3, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
+; GFX11-NEXT:    s_branch .LBB57_5
+; GFX11-NEXT:  .LBB57_3:
+; GFX11-NEXT:    s_branch .LBB57_2
+; GFX11-NEXT:  .LBB57_4:
+; GFX11-NEXT:    v_dual_mov_b32 v13, s29 :: v_dual_mov_b32 v14, s28
+; GFX11-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v16, s26
+; GFX11-NEXT:    v_dual_mov_b32 v17, s25 :: v_dual_mov_b32 v8, s24
+; GFX11-NEXT:    v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v10, s22
+; GFX11-NEXT:    v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v12, s20
+; GFX11-NEXT:    v_dual_mov_b32 v3, s19 :: v_dual_mov_b32 v4, s18
+; GFX11-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s16
+; GFX11-NEXT:    v_dual_mov_b32 v7, s3 :: v_dual_mov_b32 v0, s2
+; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s0
+; GFX11-NEXT:    v_dual_mov_b32 v18, s45 :: v_dual_mov_b32 v19, s44
+; GFX11-NEXT:    v_dual_mov_b32 v20, s43 :: v_dual_mov_b32 v21, s42
+; GFX11-NEXT:    v_dual_mov_b32 v22, s41 :: v_dual_mov_b32 v23, s40
+; GFX11-NEXT:    v_dual_mov_b32 v24, s15 :: v_dual_mov_b32 v25, s14
+; GFX11-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
+; GFX11-NEXT:    v_dual_mov_b32 v28, s11 :: v_dual_mov_b32 v29, s10
+; GFX11-NEXT:    v_dual_mov_b32 v30, s9 :: v_dual_mov_b32 v31, s7
+; GFX11-NEXT:    v_dual_mov_b32 v32, s6 :: v_dual_mov_b32 v33, s8
+; GFX11-NEXT:    v_dual_mov_b32 v34, s4 :: v_dual_mov_b32 v35, s5
+; GFX11-NEXT:  .LBB57_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v0, v35, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v2, v33, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v3, v32, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v4, v31, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v6, v29, 16, v33
+; GFX11-NEXT:    v_lshl_or_b32 v7, v28, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshl_or_b32 v8, v27, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v9, v26, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v11, v24, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v12, v23, 16, v29
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
+; GFX11-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v13, v22, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v14, v21, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v19, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v17, v18, 16, v24
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -31860,317 +29122,149 @@ define inreg <36 x i16> @bitcast_v36f16_to_v36i16_scalar(<36 x half> inreg %a, i
 ; GFX9-NEXT:    v_mov_b32_e32 v3, v19
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v36f16_to_v36i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB59_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB59_4
-; GFX11-TRUE16-NEXT:  .LBB59_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s15, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s14, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s13, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s12, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s11, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s10, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s29 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s28 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s26 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s25 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s24 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v13
-; GFX11-TRUE16-NEXT:    s_branch .LBB59_5
-; GFX11-TRUE16-NEXT:  .LBB59_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB59_2
-; GFX11-TRUE16-NEXT:  .LBB59_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s5
-; GFX11-TRUE16-NEXT:  .LBB59_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v28, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v29, 16, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v23, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v24, 16, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v35, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v27, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v18, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v22, 16, v24
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v36f16_to_v36i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s0, 16
-; GFX11-FAKE16-NEXT:    s_mov_b32 s46, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB59_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB59_4
-; GFX11-FAKE16-NEXT:  .LBB59_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s21, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s20, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s19, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s18, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s17, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s6
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v13, 0x200, s29 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v14, 0x200, s28 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v16, 0x200, s26 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v17, 0x200, s25 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v8, 0x200, s24 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v9, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v10, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v11, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v3, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v4, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v5, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v0, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v7, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
-; GFX11-FAKE16-NEXT:    s_branch .LBB59_5
-; GFX11-FAKE16-NEXT:  .LBB59_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB59_2
-; GFX11-FAKE16-NEXT:  .LBB59_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s29 :: v_dual_mov_b32 v14, s28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v16, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s25 :: v_dual_mov_b32 v8, s24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v10, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v12, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s19 :: v_dual_mov_b32 v4, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s3 :: v_dual_mov_b32 v0, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s45 :: v_dual_mov_b32 v19, s44
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s43 :: v_dual_mov_b32 v21, s42
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s41 :: v_dual_mov_b32 v23, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s15 :: v_dual_mov_b32 v25, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s11 :: v_dual_mov_b32 v29, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s9 :: v_dual_mov_b32 v31, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, s6 :: v_dual_mov_b32 v33, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, s4 :: v_dual_mov_b32 v35, s5
-; GFX11-FAKE16-NEXT:  .LBB59_5: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v35, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v33, 16, v36
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v32, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v31, 16, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v29, 16, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v28, 16, v34
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v27, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v26, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v24, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v23, 16, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v22, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v21, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v19, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v24
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v36f16_to_v36i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    s_lshr_b32 s45, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s0, 16
+; GFX11-NEXT:    s_mov_b32 s46, 0
+; GFX11-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB59_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-NEXT:    s_cbranch_vccnz .LBB59_4
+; GFX11-NEXT:  .LBB59_2: ; %cmp.true
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s21, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s20, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s19, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s18, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s17, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s6
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s29 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s28 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s26 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s25 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s24 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v19, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 16, v13
+; GFX11-NEXT:    s_branch .LBB59_5
+; GFX11-NEXT:  .LBB59_3:
+; GFX11-NEXT:    s_branch .LBB59_2
+; GFX11-NEXT:  .LBB59_4:
+; GFX11-NEXT:    v_dual_mov_b32 v13, s29 :: v_dual_mov_b32 v14, s28
+; GFX11-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v16, s26
+; GFX11-NEXT:    v_dual_mov_b32 v17, s25 :: v_dual_mov_b32 v8, s24
+; GFX11-NEXT:    v_dual_mov_b32 v9, s23 :: v_dual_mov_b32 v10, s22
+; GFX11-NEXT:    v_dual_mov_b32 v11, s21 :: v_dual_mov_b32 v12, s20
+; GFX11-NEXT:    v_dual_mov_b32 v3, s19 :: v_dual_mov_b32 v4, s18
+; GFX11-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s16
+; GFX11-NEXT:    v_dual_mov_b32 v7, s3 :: v_dual_mov_b32 v0, s2
+; GFX11-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s0
+; GFX11-NEXT:    v_dual_mov_b32 v18, s45 :: v_dual_mov_b32 v19, s44
+; GFX11-NEXT:    v_dual_mov_b32 v20, s43 :: v_dual_mov_b32 v21, s42
+; GFX11-NEXT:    v_dual_mov_b32 v22, s41 :: v_dual_mov_b32 v23, s40
+; GFX11-NEXT:    v_dual_mov_b32 v24, s15 :: v_dual_mov_b32 v25, s14
+; GFX11-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
+; GFX11-NEXT:    v_dual_mov_b32 v28, s11 :: v_dual_mov_b32 v29, s10
+; GFX11-NEXT:    v_dual_mov_b32 v30, s9 :: v_dual_mov_b32 v31, s7
+; GFX11-NEXT:    v_dual_mov_b32 v32, s6 :: v_dual_mov_b32 v33, s8
+; GFX11-NEXT:    v_dual_mov_b32 v34, s4 :: v_dual_mov_b32 v35, s5
+; GFX11-NEXT:  .LBB59_5: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v0, v35, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v1, v34, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v2, v33, 16, v36
+; GFX11-NEXT:    v_and_b32_e32 v33, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v34, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v3, v32, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v4, v31, 16, v6
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v6, v29, 16, v33
+; GFX11-NEXT:    v_lshl_or_b32 v7, v28, 16, v34
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_lshl_or_b32 v8, v27, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v9, v26, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v11, v24, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v12, v23, 16, v29
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v5, v30, 16, v5
+; GFX11-NEXT:    v_lshl_or_b32 v10, v25, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v13, v22, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v14, v21, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v15, v20, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v19, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v17, v18, 16, v24
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.640bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.640bit.ll
index cc55ba1d84df6..b8091d8256457 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.640bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.640bit.ll
@@ -3394,268 +3394,134 @@ define inreg <40 x i16> @bitcast_v20i32_to_v40i16_scalar(<20 x i32> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr12
 ; GFX9-NEXT:    s_branch .LBB13_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v20i32_to_v40i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-TRUE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB13_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s22, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s23, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB13_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB13_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v20i32_to_v40i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s58, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-FAKE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB13_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s25, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s26, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s27, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s28, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s29, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB13_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB13_2
+; GFX11-LABEL: bitcast_v20i32_to_v40i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-NEXT:    s_mov_b32 s58, 0
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB13_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s6, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
+; GFX11-NEXT:    s_cbranch_vccnz .LBB13_3
+; GFX11-NEXT:  .LBB13_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s6, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s0, 16
+; GFX11-NEXT:  .LBB13_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s25, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s26, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s27, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s28, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s29, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
+; GFX11-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
+; GFX11-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
+; GFX11-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
+; GFX11-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB13_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB13_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -5049,144 +4915,88 @@ define inreg <20 x i32> @bitcast_v40i16_to_v20i32_scalar(<40 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v40i16_to_v20i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v1 :: v_dual_mov_b32 v33, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v1
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s1, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s2, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s25, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s26, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB15_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v13, s76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s7 :: v_dual_mov_b32 v5, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s9 :: v_dual_mov_b32 v7, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s11 :: v_dual_mov_b32 v9, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v11, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s15 :: v_dual_mov_b32 v13, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s17 :: v_dual_mov_b32 v15, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s1 :: v_dual_mov_b32 v17, s2
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB15_3
 ; GFX11-TRUE16-NEXT:  .LBB15_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:  .LBB15_3: ; %end
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -6768,268 +6578,134 @@ define inreg <40 x half> @bitcast_v20i32_to_v40f16_scalar(<20 x i32> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr12
 ; GFX9-NEXT:    s_branch .LBB17_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v20i32_to_v40f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-TRUE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB17_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s22, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s23, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB17_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB17_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v20i32_to_v40f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s58, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-FAKE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB17_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s25, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s26, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s27, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s28, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s29, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB17_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB17_2
+; GFX11-LABEL: bitcast_v20i32_to_v40f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-NEXT:    s_mov_b32 s58, 0
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB17_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s6, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
+; GFX11-NEXT:    s_cbranch_vccnz .LBB17_3
+; GFX11-NEXT:  .LBB17_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s6, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s0, 16
+; GFX11-NEXT:  .LBB17_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s25, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s26, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s27, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s28, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s29, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
+; GFX11-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
+; GFX11-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
+; GFX11-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
+; GFX11-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB17_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB17_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -8668,144 +8344,88 @@ define inreg <20 x i32> @bitcast_v40f16_to_v20i32_scalar(<40 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v40f16_to_v20i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v1 :: v_dual_mov_b32 v33, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v1
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s1, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s2, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s25, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s26, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB19_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v13, s76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s7 :: v_dual_mov_b32 v5, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s9 :: v_dual_mov_b32 v7, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s11 :: v_dual_mov_b32 v9, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v11, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s15 :: v_dual_mov_b32 v13, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s17 :: v_dual_mov_b32 v15, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s1 :: v_dual_mov_b32 v17, s2
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB19_3
 ; GFX11-TRUE16-NEXT:  .LBB19_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:  .LBB19_3: ; %end
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -11482,279 +11102,142 @@ define inreg <40 x i16> @bitcast_v20f32_to_v40i16_scalar(<20 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr26
 ; GFX9-NEXT:    s_branch .LBB29_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v20f32_to_v40i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v19, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s2 :: v_dual_mov_b32 v3, s16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s3 :: v_dual_mov_b32 v9, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s20 :: v_dual_mov_b32 v5, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s22 :: v_dual_mov_b32 v13, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v11, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s26 :: v_dual_mov_b32 v17, s28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s27 :: v_dual_mov_b32 v15, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-TRUE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-TRUE16-NEXT:  .LBB29_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v36, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v33, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v32, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v28, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v27, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v38, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v37, 16, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v29, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v26, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v24, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v23, 16, v0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v20
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB29_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr22_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB29_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v20f32_to_v40i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v19, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s2 :: v_dual_mov_b32 v3, s16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s3 :: v_dual_mov_b32 v9, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s20 :: v_dual_mov_b32 v5, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s22 :: v_dual_mov_b32 v13, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v11, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-FAKE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-FAKE16-NEXT:  .LBB29_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v36, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v35, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v33, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v32, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v28, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v27, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v38, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v37, 16, v48
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v29, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v25, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v24, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v20
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB29_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr21
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr22
-; GFX11-FAKE16-NEXT:    s_branch .LBB29_2
+; GFX11-LABEL: bitcast_v20f32_to_v40i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v19, s1
+; GFX11-NEXT:    v_dual_mov_b32 v18, s2 :: v_dual_mov_b32 v3, s16
+; GFX11-NEXT:    v_dual_mov_b32 v4, s3 :: v_dual_mov_b32 v9, s17
+; GFX11-NEXT:    v_dual_mov_b32 v8, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v6, s20 :: v_dual_mov_b32 v5, s21
+; GFX11-NEXT:    v_dual_mov_b32 v14, s22 :: v_dual_mov_b32 v13, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v11, s25
+; GFX11-NEXT:    v_dual_mov_b32 v10, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB29_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB29_3
+; GFX11-NEXT:  .LBB29_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
+; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v10, 1.0, v10
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
+; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v6, 1.0, v6
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v4, 1.0, v4
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
+; GFX11-NEXT:  .LBB29_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v21, v21, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v2, v2, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v6
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v5, v36, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v6, v35, 16, v8
+; GFX11-NEXT:    v_lshl_or_b32 v8, v33, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v9, v32, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v10, v31, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v13, v28, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v14, v27, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v19, v22, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v21
+; GFX11-NEXT:    v_lshl_or_b32 v20, v39, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v3, v38, 16, v4
+; GFX11-NEXT:    v_lshl_or_b32 v4, v37, 16, v48
+; GFX11-NEXT:    v_lshl_or_b32 v7, v34, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v12, v29, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v15, v26, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v25, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v24, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v23, 16, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, v20
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB29_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr21
+; GFX11-NEXT:    ; implicit-def: $vgpr2
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    ; implicit-def: $vgpr25
+; GFX11-NEXT:    ; implicit-def: $vgpr24
+; GFX11-NEXT:    ; implicit-def: $vgpr23
+; GFX11-NEXT:    ; implicit-def: $vgpr22
+; GFX11-NEXT:    s_branch .LBB29_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -13148,144 +12631,88 @@ define inreg <20 x float> @bitcast_v40i16_to_v20f32_scalar(<40 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v40i16_to_v20f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v1 :: v_dual_mov_b32 v33, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v1
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s1, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s2, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s25, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s26, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB31_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v13, s76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s7 :: v_dual_mov_b32 v5, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s9 :: v_dual_mov_b32 v7, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s11 :: v_dual_mov_b32 v9, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v11, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s15 :: v_dual_mov_b32 v13, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s17 :: v_dual_mov_b32 v15, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s1 :: v_dual_mov_b32 v17, s2
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB31_3
 ; GFX11-TRUE16-NEXT:  .LBB31_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:  .LBB31_3: ; %end
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -14844,279 +14271,142 @@ define inreg <40 x half> @bitcast_v20f32_to_v40f16_scalar(<20 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr26
 ; GFX9-NEXT:    s_branch .LBB33_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v20f32_to_v40f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v19, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s2 :: v_dual_mov_b32 v3, s16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s3 :: v_dual_mov_b32 v9, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s20 :: v_dual_mov_b32 v5, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s22 :: v_dual_mov_b32 v13, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v11, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s26 :: v_dual_mov_b32 v17, s28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s27 :: v_dual_mov_b32 v15, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-TRUE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-TRUE16-NEXT:  .LBB33_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v36, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v33, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v32, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v28, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v27, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v38, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v37, 16, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v29, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v26, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v24, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v23, 16, v0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v20
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB33_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr22_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB33_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v20f32_to_v40f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v19, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s2 :: v_dual_mov_b32 v3, s16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s3 :: v_dual_mov_b32 v9, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s20 :: v_dual_mov_b32 v5, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s22 :: v_dual_mov_b32 v13, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v11, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-FAKE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-FAKE16-NEXT:  .LBB33_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v36, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v35, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v33, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v32, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v28, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v27, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v38, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v37, 16, v48
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v29, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v25, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v24, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v20
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB33_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr21
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr22
-; GFX11-FAKE16-NEXT:    s_branch .LBB33_2
+; GFX11-LABEL: bitcast_v20f32_to_v40f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v19, s1
+; GFX11-NEXT:    v_dual_mov_b32 v18, s2 :: v_dual_mov_b32 v3, s16
+; GFX11-NEXT:    v_dual_mov_b32 v4, s3 :: v_dual_mov_b32 v9, s17
+; GFX11-NEXT:    v_dual_mov_b32 v8, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v6, s20 :: v_dual_mov_b32 v5, s21
+; GFX11-NEXT:    v_dual_mov_b32 v14, s22 :: v_dual_mov_b32 v13, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v11, s25
+; GFX11-NEXT:    v_dual_mov_b32 v10, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB33_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB33_3
+; GFX11-NEXT:  .LBB33_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
+; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v10, 1.0, v10
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
+; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v6, 1.0, v6
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v4, 1.0, v4
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
+; GFX11-NEXT:  .LBB33_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v21, v21, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v2, v2, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v6
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v5, v36, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v6, v35, 16, v8
+; GFX11-NEXT:    v_lshl_or_b32 v8, v33, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v9, v32, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v48, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v10, v31, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v13, v28, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v14, v27, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v19, v22, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v21
+; GFX11-NEXT:    v_lshl_or_b32 v20, v39, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v3, v38, 16, v4
+; GFX11-NEXT:    v_lshl_or_b32 v4, v37, 16, v48
+; GFX11-NEXT:    v_lshl_or_b32 v7, v34, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v12, v29, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v15, v26, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v25, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v24, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v23, 16, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, v20
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB33_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr21
+; GFX11-NEXT:    ; implicit-def: $vgpr2
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    ; implicit-def: $vgpr25
+; GFX11-NEXT:    ; implicit-def: $vgpr24
+; GFX11-NEXT:    ; implicit-def: $vgpr23
+; GFX11-NEXT:    ; implicit-def: $vgpr22
+; GFX11-NEXT:    s_branch .LBB33_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -16755,144 +16045,88 @@ define inreg <20 x float> @bitcast_v40f16_to_v20f32_scalar(<40 x half> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v40f16_to_v20f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v1 :: v_dual_mov_b32 v33, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v1
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s1, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s2, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s25, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s26, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB35_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v13, s76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s7 :: v_dual_mov_b32 v5, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s9 :: v_dual_mov_b32 v7, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s11 :: v_dual_mov_b32 v9, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v11, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s15 :: v_dual_mov_b32 v13, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s17 :: v_dual_mov_b32 v15, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s1 :: v_dual_mov_b32 v17, s2
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB35_3
 ; GFX11-TRUE16-NEXT:  .LBB35_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:  .LBB35_3: ; %end
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -18904,268 +18138,134 @@ define inreg <40 x i16> @bitcast_v10i64_to_v40i16_scalar(<10 x i64> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr12
 ; GFX9-NEXT:    s_branch .LBB41_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v10i64_to_v40i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-TRUE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB41_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s22, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s23, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB41_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB41_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v10i64_to_v40i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s58, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-FAKE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB41_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s25, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s26, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s27, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s28, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s29, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB41_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB41_2
+; GFX11-LABEL: bitcast_v10i64_to_v40i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-NEXT:    s_mov_b32 s58, 0
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB41_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s6, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
+; GFX11-NEXT:    s_cbranch_vccnz .LBB41_3
+; GFX11-NEXT:  .LBB41_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s5, s5, 3
+; GFX11-NEXT:    s_addc_u32 s4, s4, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s6, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s0, 16
+; GFX11-NEXT:  .LBB41_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s25, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s26, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s27, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s28, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s29, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
+; GFX11-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
+; GFX11-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
+; GFX11-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
+; GFX11-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB41_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB41_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -20559,144 +19659,88 @@ define inreg <10 x i64> @bitcast_v40i16_to_v10i64_scalar(<40 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v40i16_to_v10i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v1 :: v_dual_mov_b32 v33, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v1
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s1, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s2, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s25, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s26, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB43_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v13, s76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s7 :: v_dual_mov_b32 v5, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s9 :: v_dual_mov_b32 v7, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s11 :: v_dual_mov_b32 v9, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v11, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s15 :: v_dual_mov_b32 v13, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s17 :: v_dual_mov_b32 v15, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s1 :: v_dual_mov_b32 v17, s2
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB43_3
 ; GFX11-TRUE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:  .LBB43_3: ; %end
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -22288,268 +21332,134 @@ define inreg <40 x half> @bitcast_v10i64_to_v40f16_scalar(<10 x i64> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr12
 ; GFX9-NEXT:    s_branch .LBB45_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v10i64_to_v40f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-TRUE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB45_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s23, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s22, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s23, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB45_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB45_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v10i64_to_v40f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s58, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-FAKE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB45_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s25, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s26, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s27, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s28, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s29, s8
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB45_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB45_2
+; GFX11-LABEL: bitcast_v10i64_to_v40f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-NEXT:    s_mov_b32 s58, 0
+; GFX11-NEXT:    s_and_b32 s6, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB45_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s6, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s58
+; GFX11-NEXT:    s_cbranch_vccnz .LBB45_3
+; GFX11-NEXT:  .LBB45_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s5, s5, 3
+; GFX11-NEXT:    s_addc_u32 s4, s4, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s6, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s0, 16
+; GFX11-NEXT:  .LBB45_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s22, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s25, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s26, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s27, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s28, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s29, s8
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s7
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s6
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s15 :: v_dual_mov_b32 v11, s14
+; GFX11-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s12
+; GFX11-NEXT:    v_dual_mov_b32 v14, s11 :: v_dual_mov_b32 v15, s10
+; GFX11-NEXT:    v_dual_mov_b32 v16, s9 :: v_dual_mov_b32 v17, s8
+; GFX11-NEXT:    v_dual_mov_b32 v18, s5 :: v_dual_mov_b32 v19, s4
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB45_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB45_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -24188,144 +23098,88 @@ define inreg <10 x i64> @bitcast_v40f16_to_v10i64_scalar(<40 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v40f16_to_v10i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v1 :: v_dual_mov_b32 v33, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v1
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s1, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s2, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s25, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s26, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB47_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v13, s76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s7 :: v_dual_mov_b32 v5, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s9 :: v_dual_mov_b32 v7, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s11 :: v_dual_mov_b32 v9, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v11, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s15 :: v_dual_mov_b32 v13, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s17 :: v_dual_mov_b32 v15, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s1 :: v_dual_mov_b32 v17, s2
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB47_3
 ; GFX11-TRUE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:  .LBB47_3: ; %end
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -25532,278 +24386,142 @@ define inreg <40 x i16> @bitcast_v10f64_to_v40i16_scalar(<10 x double> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr26
 ; GFX9-NEXT:    s_branch .LBB49_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v10f64_to_v40i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v19, s0 :: v_dual_mov_b32 v20, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v21, s16 :: v_dual_mov_b32 v22, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s20 :: v_dual_mov_b32 v6, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v13, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v11, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v18, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-TRUE16-NEXT:  .LBB49_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v35, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v36, 16, v49
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v33, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v32, 16, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v14, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v27, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v26, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v25, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v24, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v23, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v20 :: v_dual_mov_b32 v1, v21
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v10f64_to_v40i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v21, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v19, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s20 :: v_dual_mov_b32 v6, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v13, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v11, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-FAKE16-NEXT:  .LBB49_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v36, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v33, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v32, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v35, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v28, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v29, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v27, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v25, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v24, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v20
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr22
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
+; GFX11-LABEL: bitcast_v10f64_to_v40i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v21, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v19, s17
+; GFX11-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
+; GFX11-NEXT:    v_dual_mov_b32 v5, s20 :: v_dual_mov_b32 v6, s21
+; GFX11-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v13, s23
+; GFX11-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v11, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_3
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
+; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
+; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
+; GFX11-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
+; GFX11-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
+; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
+; GFX11-NEXT:  .LBB49_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v2, v2, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v3, v37, 16, v4
+; GFX11-NEXT:    v_lshl_or_b32 v4, v36, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v21, v38, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v6, v34, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v7, v33, 16, v8
+; GFX11-NEXT:    v_lshl_or_b32 v8, v32, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v5, v35, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v13, v28, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v19, v22, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v21
+; GFX11-NEXT:    v_lshl_or_b32 v20, v39, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v9, v9, 16, v36
+; GFX11-NEXT:    v_lshl_or_b32 v10, v31, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v29, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v14, v27, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v15, v26, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v25, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v24, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v18, v23, 16, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, v20
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr2
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr9
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    ; implicit-def: $vgpr25
+; GFX11-NEXT:    ; implicit-def: $vgpr24
+; GFX11-NEXT:    ; implicit-def: $vgpr23
+; GFX11-NEXT:    ; implicit-def: $vgpr22
+; GFX11-NEXT:    s_branch .LBB49_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -27197,144 +25915,88 @@ define inreg <10 x double> @bitcast_v40i16_to_v10f64_scalar(<40 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v40i16_to_v10f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v1 :: v_dual_mov_b32 v33, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v1
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s1, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s2, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s25, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s26, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB51_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v13, s76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s7 :: v_dual_mov_b32 v5, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s9 :: v_dual_mov_b32 v7, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s11 :: v_dual_mov_b32 v9, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v11, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s15 :: v_dual_mov_b32 v13, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s17 :: v_dual_mov_b32 v15, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s1 :: v_dual_mov_b32 v17, s2
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB51_3
 ; GFX11-TRUE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:  .LBB51_3: ; %end
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -28826,278 +27488,142 @@ define inreg <40 x half> @bitcast_v10f64_to_v40f16_scalar(<10 x double> inreg %a
 ; GFX9-NEXT:    ; implicit-def: $vgpr26
 ; GFX9-NEXT:    s_branch .LBB53_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v10f64_to_v40f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v19, s0 :: v_dual_mov_b32 v20, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v21, s16 :: v_dual_mov_b32 v22, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s20 :: v_dual_mov_b32 v6, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v13, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v11, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v18, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-TRUE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-TRUE16-NEXT:  .LBB53_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v35, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v36, 16, v49
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v33, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v32, 16, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v14, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v27, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v26, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v25, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v24, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v23, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v20 :: v_dual_mov_b32 v1, v21
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB53_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB53_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v10f64_to_v40f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v21, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v19, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s20 :: v_dual_mov_b32 v6, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v13, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v11, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-FAKE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
-; GFX11-FAKE16-NEXT:  .LBB53_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v36, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v33, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v32, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v35, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v28, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v29, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v27, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v26, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v25, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v24, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v23, 16, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v20
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB53_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr22
-; GFX11-FAKE16-NEXT:    s_branch .LBB53_2
+; GFX11-LABEL: bitcast_v10f64_to_v40f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v21, s1
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s3
+; GFX11-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v19, s17
+; GFX11-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
+; GFX11-NEXT:    v_dual_mov_b32 v5, s20 :: v_dual_mov_b32 v6, s21
+; GFX11-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v13, s23
+; GFX11-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v11, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB53_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB53_3
+; GFX11-NEXT:  .LBB53_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
+; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
+; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
+; GFX11-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
+; GFX11-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
+; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v20
+; GFX11-NEXT:  .LBB53_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v2, v2, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v3, v37, 16, v4
+; GFX11-NEXT:    v_lshl_or_b32 v4, v36, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v21, v38, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v36, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v6, v34, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v7, v33, 16, v8
+; GFX11-NEXT:    v_lshl_or_b32 v8, v32, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v5, v35, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v11, v30, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v13, v28, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v19, v22, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v21
+; GFX11-NEXT:    v_lshl_or_b32 v20, v39, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v9, v9, 16, v36
+; GFX11-NEXT:    v_lshl_or_b32 v10, v31, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v29, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v14, v27, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v15, v26, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v25, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v24, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v18, v23, 16, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, v20
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB53_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr2
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr9
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    ; implicit-def: $vgpr25
+; GFX11-NEXT:    ; implicit-def: $vgpr24
+; GFX11-NEXT:    ; implicit-def: $vgpr23
+; GFX11-NEXT:    ; implicit-def: $vgpr22
+; GFX11-NEXT:    s_branch .LBB53_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -30736,144 +29262,88 @@ define inreg <10 x double> @bitcast_v40f16_to_v10f64_scalar(<40 x half> inreg %a
 ; GFX11-TRUE16-LABEL: bitcast_v40f16_to_v10f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v1 :: v_dual_mov_b32 v33, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v1
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s1, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s2, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s25, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s26, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB55_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v13, s76
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s7 :: v_dual_mov_b32 v5, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s9 :: v_dual_mov_b32 v7, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s11 :: v_dual_mov_b32 v9, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v11, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s15 :: v_dual_mov_b32 v13, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s17 :: v_dual_mov_b32 v15, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s1 :: v_dual_mov_b32 v17, s2
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB55_3
 ; GFX11-TRUE16-NEXT:  .LBB55_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v32, 16, v34
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:  .LBB55_3: ; %end
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
@@ -32591,8 +31061,12 @@ define inreg <40 x half> @bitcast_v40i16_to_v40f16_scalar(<40 x i16> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v40i16_to_v40f16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v19.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
@@ -32606,10 +31080,10 @@ define inreg <40 x half> @bitcast_v40i16_to_v40f16_scalar(<40 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s18, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB57_3
@@ -32617,161 +31091,131 @@ define inreg <40 x half> @bitcast_v40i16_to_v40f16_scalar(<40 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB57_4
 ; GFX11-TRUE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s15, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v3, 16, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s15, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s14, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s13, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s12, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s16
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s11, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s10, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s21, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s20, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s19, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s18, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s29, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s28, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s27, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s26, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s25, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s24, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s10, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s8, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, s1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s2, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v21
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v20
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v38.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v39.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_5
 ; GFX11-TRUE16-NEXT:  .LBB57_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_2
 ; GFX11-TRUE16-NEXT:  .LBB57_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s29 :: v_dual_mov_b32 v16, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s27 :: v_dual_mov_b32 v10, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s25 :: v_dual_mov_b32 v12, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s23 :: v_dual_mov_b32 v14, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s21 :: v_dual_mov_b32 v6, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v8, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s17 :: v_dual_mov_b32 v2, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v4, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s1 :: v_dual_mov_b32 v21, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v23, s44
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s43 :: v_dual_mov_b32 v25, s42
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v27, s40
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s13 :: v_dual_mov_b32 v31, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s11 :: v_dual_mov_b32 v33, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s8 :: v_dual_mov_b32 v35, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, s6 :: v_dual_mov_b32 v37, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s4 :: v_dual_mov_b32 v39, s9
 ; GFX11-TRUE16-NEXT:  .LBB57_5: ; %end
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v20
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v37, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v20
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v36, 16, v3
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v33, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v31, 16, v35
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v21
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v34, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v30, 16, v36
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v28, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v29, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v23, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v24, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v25, 16, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v29, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v25, 16, v31
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v35, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v36, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v37, 16, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v30, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v27, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v22, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v24, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v25, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v32, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v28, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v27, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v26, 16, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v24, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v23, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v22, 16, v25
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v20 :: v_dual_mov_b32 v1, v21
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -34410,8 +32854,12 @@ define inreg <40 x i16> @bitcast_v40f16_to_v40i16_scalar(<40 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v40f16_to_v40i16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v19.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
@@ -34425,10 +32873,10 @@ define inreg <40 x i16> @bitcast_v40f16_to_v40i16_scalar(<40 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s18, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB59_3
@@ -34436,161 +32884,131 @@ define inreg <40 x i16> @bitcast_v40f16_to_v40i16_scalar(<40 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB59_4
 ; GFX11-TRUE16-NEXT:  .LBB59_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v3, 16, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s15, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s14, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s13, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s12, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s16
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s11, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s10, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s21, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s20, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s19, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s18, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s16, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s29 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s28 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s27 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s26 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s25 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s24 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s10 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s8 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, s1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s2 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v21
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v20
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v38.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v39.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_5
 ; GFX11-TRUE16-NEXT:  .LBB59_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_2
 ; GFX11-TRUE16-NEXT:  .LBB59_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s29 :: v_dual_mov_b32 v16, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s27 :: v_dual_mov_b32 v10, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s25 :: v_dual_mov_b32 v12, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s23 :: v_dual_mov_b32 v14, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s21 :: v_dual_mov_b32 v6, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v8, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s17 :: v_dual_mov_b32 v2, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v4, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s1 :: v_dual_mov_b32 v21, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s45 :: v_dual_mov_b32 v23, s44
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s43 :: v_dual_mov_b32 v25, s42
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s41 :: v_dual_mov_b32 v27, s40
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s13 :: v_dual_mov_b32 v31, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s11 :: v_dual_mov_b32 v33, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s8 :: v_dual_mov_b32 v35, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, s6 :: v_dual_mov_b32 v37, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s4 :: v_dual_mov_b32 v39, s9
 ; GFX11-TRUE16-NEXT:  .LBB59_5: ; %end
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v20
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v37, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v20
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v36, 16, v3
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v33, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v31, 16, v35
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v21
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v34, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v30, 16, v36
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v38
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v28, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v29, 16, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v23, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v24, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v25, 16, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v29, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v25, 16, v31
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v35, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v36, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v37, 16, v39
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v30, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v26, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v27, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v22, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v24, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v25, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v32, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v28, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v27, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v26, 16, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v24, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v23, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v22, 16, v25
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v20 :: v_dual_mov_b32 v1, v21
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll
index b73870977c429..a40ee1698b8e0 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll
@@ -2411,128 +2411,66 @@ define inreg i64 @bitcast_v4bf16_to_i64_scalar(<4 x bfloat> inreg %a, i32 inreg
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4bf16_to_i64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB23_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB23_4
-; GFX11-TRUE16-NEXT:  .LBB23_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s1, 16
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v4, v10 :: v_dual_add_nc_u32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v5, v6 :: v_dual_and_b32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB23_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB23_2
-; GFX11-TRUE16-NEXT:  .LBB23_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4bf16_to_i64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB23_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB23_4
-; GFX11-FAKE16-NEXT:  .LBB23_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s2, 0, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s1, 16
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB23_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB23_2
-; GFX11-FAKE16-NEXT:  .LBB23_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4bf16_to_i64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s2, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB23_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
+; GFX11-NEXT:    s_cbranch_vccnz .LBB23_4
+; GFX11-NEXT:  .LBB23_2: ; %cmp.true
+; GFX11-NEXT:    s_pack_lh_b32_b16 s2, 0, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s1, 16
+; GFX11-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v3, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v2
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB23_3:
+; GFX11-NEXT:    s_branch .LBB23_2
+; GFX11-NEXT:  .LBB23_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -2885,88 +2823,45 @@ define inreg <8 x i8> @bitcast_i64_to_v8i8_scalar(i64 inreg %a, i32 inreg %b) {
 ; GFX9-NEXT:    ; implicit-def: $sgpr5
 ; GFX9-NEXT:    s_branch .LBB25_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_i64_to_v8i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB25_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB25_3
-; GFX11-TRUE16-NEXT:  .LBB25_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB25_3: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s3
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB25_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB25_2
-;
-; GFX11-FAKE16-LABEL: bitcast_i64_to_v8i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB25_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB25_3
-; GFX11-FAKE16-NEXT:  .LBB25_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB25_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s4 :: v_dual_mov_b32 v7, s3
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB25_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    s_branch .LBB25_2
+; GFX11-LABEL: bitcast_i64_to_v8i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s8, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB25_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s3, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s6, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
+; GFX11-NEXT:    s_cbranch_vccnz .LBB25_3
+; GFX11-NEXT:  .LBB25_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s6, s0, 16
+; GFX11-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s3, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s7, s0, 8
+; GFX11-NEXT:  .LBB25_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s7
+; GFX11-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s2
+; GFX11-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT:    v_dual_mov_b32 v6, s4 :: v_dual_mov_b32 v7, s3
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB25_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    s_branch .LBB25_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -5663,128 +5558,66 @@ define inreg double @bitcast_v4bf16_to_f64_scalar(<4 x bfloat> inreg %a, i32 inr
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4bf16_to_f64_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB47_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB47_4
-; GFX11-TRUE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s1, 16
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v4, v10 :: v_dual_add_nc_u32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v5, v6 :: v_dual_and_b32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB47_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB47_2
-; GFX11-TRUE16-NEXT:  .LBB47_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4bf16_to_f64_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB47_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB47_4
-; GFX11-FAKE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s2, 0, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s1, 16
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB47_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB47_2
-; GFX11-FAKE16-NEXT:  .LBB47_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4bf16_to_f64_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s2, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB47_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
+; GFX11-NEXT:    s_cbranch_vccnz .LBB47_4
+; GFX11-NEXT:  .LBB47_2: ; %cmp.true
+; GFX11-NEXT:    s_pack_lh_b32_b16 s2, 0, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s1, 16
+; GFX11-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v3, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v2
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB47_3:
+; GFX11-NEXT:    s_branch .LBB47_2
+; GFX11-NEXT:  .LBB47_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -6138,96 +5971,49 @@ define inreg <8 x i8> @bitcast_f64_to_v8i8_scalar(double inreg %a, i32 inreg %b)
 ; GFX9-NEXT:    v_mov_b32_e32 v4, v9
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_f64_to_v8i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[3:4], s[0:1], 1.0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_5
-; GFX11-TRUE16-NEXT:  .LBB49_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s6
-; GFX11-TRUE16-NEXT:  .LBB49_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_f64_to_v8i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_4
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[8:9], s[0:1], 1.0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[8:9]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v8
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_5
-; GFX11-FAKE16-NEXT:  .LBB49_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s0 :: v_dual_mov_b32 v9, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s8 :: v_dual_mov_b32 v1, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v6, s5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v7, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v5, s3
-; GFX11-FAKE16-NEXT:  .LBB49_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v9
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_f64_to_v8i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s6, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s3, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s8, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_4
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[8:9], s[0:1], 1.0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[8:9]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v8
+; GFX11-NEXT:    s_branch .LBB49_5
+; GFX11-NEXT:  .LBB49_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB49_2
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    v_dual_mov_b32 v8, s0 :: v_dual_mov_b32 v9, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s8 :: v_dual_mov_b32 v1, s7
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v6, s5
+; GFX11-NEXT:    v_mov_b32_e32 v7, s6
+; GFX11-NEXT:    v_mov_b32_e32 v5, s3
+; GFX11-NEXT:  .LBB49_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v8
+; GFX11-NEXT:    v_mov_b32_e32 v4, v9
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -8631,128 +8417,66 @@ define inreg <2 x i32> @bitcast_v4bf16_to_v2i32_scalar(<4 x bfloat> inreg %a, i3
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4bf16_to_v2i32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB67_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB67_4
-; GFX11-TRUE16-NEXT:  .LBB67_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s1, 16
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v4, v10 :: v_dual_add_nc_u32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v5, v6 :: v_dual_and_b32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB67_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB67_2
-; GFX11-TRUE16-NEXT:  .LBB67_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4bf16_to_v2i32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB67_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB67_4
-; GFX11-FAKE16-NEXT:  .LBB67_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s2, 0, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s1, 16
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB67_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB67_2
-; GFX11-FAKE16-NEXT:  .LBB67_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4bf16_to_v2i32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s2, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB67_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
+; GFX11-NEXT:    s_cbranch_vccnz .LBB67_4
+; GFX11-NEXT:  .LBB67_2: ; %cmp.true
+; GFX11-NEXT:    s_pack_lh_b32_b16 s2, 0, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s1, 16
+; GFX11-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v3, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v2
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB67_3:
+; GFX11-NEXT:    s_branch .LBB67_2
+; GFX11-NEXT:  .LBB67_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -9103,88 +8827,45 @@ define inreg <8 x i8> @bitcast_v2i32_to_v8i8_scalar(<2 x i32> inreg %a, i32 inre
 ; GFX9-NEXT:    ; implicit-def: $sgpr5
 ; GFX9-NEXT:    s_branch .LBB69_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v2i32_to_v8i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB69_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB69_3
-; GFX11-TRUE16-NEXT:  .LBB69_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s0, 8
-; GFX11-TRUE16-NEXT:  .LBB69_3: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s3
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB69_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB69_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v2i32_to_v8i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB69_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB69_3
-; GFX11-FAKE16-NEXT:  .LBB69_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s4, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s0, 8
-; GFX11-FAKE16-NEXT:  .LBB69_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s4 :: v_dual_mov_b32 v7, s3
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB69_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    s_branch .LBB69_2
+; GFX11-LABEL: bitcast_v2i32_to_v8i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s8, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB69_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s3, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s6, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
+; GFX11-NEXT:    s_cbranch_vccnz .LBB69_3
+; GFX11-NEXT:  .LBB69_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s3, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s5, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s6, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s0, 8
+; GFX11-NEXT:  .LBB69_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s7
+; GFX11-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s2
+; GFX11-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s5
+; GFX11-NEXT:    v_dual_mov_b32 v6, s4 :: v_dual_mov_b32 v7, s3
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB69_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    s_branch .LBB69_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -11268,128 +10949,66 @@ define inreg <2 x float> @bitcast_v4bf16_to_v2f32_scalar(<4 x bfloat> inreg %a,
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4bf16_to_v2f32_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB83_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB83_4
-; GFX11-TRUE16-NEXT:  .LBB83_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s1, 16
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v4, v10 :: v_dual_add_nc_u32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v5, v6 :: v_dual_and_b32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB83_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB83_2
-; GFX11-TRUE16-NEXT:  .LBB83_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4bf16_to_v2f32_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB83_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB83_4
-; GFX11-FAKE16-NEXT:  .LBB83_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s2, 0, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s1, 16
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB83_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB83_2
-; GFX11-FAKE16-NEXT:  .LBB83_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4bf16_to_v2f32_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s2, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB83_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
+; GFX11-NEXT:    s_cbranch_vccnz .LBB83_4
+; GFX11-NEXT:  .LBB83_2: ; %cmp.true
+; GFX11-NEXT:    s_pack_lh_b32_b16 s2, 0, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s1, 16
+; GFX11-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v3, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v2
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
+; GFX11-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB83_3:
+; GFX11-NEXT:    s_branch .LBB83_2
+; GFX11-NEXT:  .LBB83_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -11746,99 +11365,50 @@ define inreg <8 x i8> @bitcast_v2f32_to_v8i8_scalar(<2 x float> inreg %a, i32 in
 ; GFX9-NEXT:    v_mov_b32_e32 v4, v9
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v2f32_to_v8i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB85_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB85_4
-; GFX11-TRUE16-NEXT:  .LBB85_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, s1, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, s0, 1.0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB85_5
-; GFX11-TRUE16-NEXT:  .LBB85_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB85_2
-; GFX11-TRUE16-NEXT:  .LBB85_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:  .LBB85_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v2f32_to_v8i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB85_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB85_4
-; GFX11-FAKE16-NEXT:  .LBB85_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, s1, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, s0, 1.0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[8:9]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v8
-; GFX11-FAKE16-NEXT:    s_branch .LBB85_5
-; GFX11-FAKE16-NEXT:  .LBB85_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    s_branch .LBB85_2
-; GFX11-FAKE16-NEXT:  .LBB85_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s0 :: v_dual_mov_b32 v9, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s8 :: v_dual_mov_b32 v2, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s6 :: v_dual_mov_b32 v6, s5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v7, s3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s2
-; GFX11-FAKE16-NEXT:  .LBB85_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v9
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v2f32_to_v8i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB85_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s3, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s7, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB85_4
+; GFX11-NEXT:  .LBB85_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f32_e64 v9, s1, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v8, s0, 1.0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v9
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[8:9]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v8
+; GFX11-NEXT:    s_branch .LBB85_5
+; GFX11-NEXT:  .LBB85_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    s_branch .LBB85_2
+; GFX11-NEXT:  .LBB85_4:
+; GFX11-NEXT:    v_dual_mov_b32 v8, s0 :: v_dual_mov_b32 v9, s1
+; GFX11-NEXT:    v_dual_mov_b32 v1, s8 :: v_dual_mov_b32 v2, s7
+; GFX11-NEXT:    v_dual_mov_b32 v5, s6 :: v_dual_mov_b32 v6, s5
+; GFX11-NEXT:    v_mov_b32_e32 v7, s3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s2
+; GFX11-NEXT:  .LBB85_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v8
+; GFX11-NEXT:    v_mov_b32_e32 v4, v9
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -13555,116 +13125,60 @@ define inreg <4 x i16> @bitcast_v4bf16_to_v4i16_scalar(<4 x bfloat> inreg %a, i3
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4bf16_to_v4i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB95_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB95_4
-; GFX11-TRUE16-NEXT:  .LBB95_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s2, 0, s1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_and_or_b32 v1, 0xffff0000, v3, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_and_or_b32 v0, 0xffff0000, v0, v2
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB95_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB95_2
-; GFX11-TRUE16-NEXT:  .LBB95_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4bf16_to_v4i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB95_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB95_4
-; GFX11-FAKE16-NEXT:  .LBB95_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s2, 0, s1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s0, 16
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_and_or_b32 v1, 0xffff0000, v3, v1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_and_or_b32 v0, 0xffff0000, v0, v2
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB95_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB95_2
-; GFX11-FAKE16-NEXT:  .LBB95_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4bf16_to_v4i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s2, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB95_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
+; GFX11-NEXT:    s_cbranch_vccnz .LBB95_4
+; GFX11-NEXT:  .LBB95_2: ; %cmp.true
+; GFX11-NEXT:    s_pack_lh_b32_b16 s2, 0, s1
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
+; GFX11-NEXT:    s_lshl_b32 s1, s0, 16
+; GFX11-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s1
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v0
+; GFX11-NEXT:    v_bfe_u32 v3, v0, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_and_or_b32 v1, 0xffff0000, v3, v1
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v4, v8, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_and_or_b32 v0, 0xffff0000, v0, v2
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB95_3:
+; GFX11-NEXT:    s_branch .LBB95_2
+; GFX11-NEXT:  .LBB95_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -14073,99 +13587,50 @@ define inreg <8 x i8> @bitcast_v4i16_to_v8i8_scalar(<4 x i16> inreg %a, i32 inre
 ; GFX9-NEXT:    v_mov_b32_e32 v4, v9
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4i16_to_v8i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB97_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB97_4
-; GFX11-TRUE16-NEXT:  .LBB97_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB97_5
-; GFX11-TRUE16-NEXT:  .LBB97_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB97_2
-; GFX11-TRUE16-NEXT:  .LBB97_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:  .LBB97_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4i16_to_v8i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB97_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB97_4
-; GFX11-FAKE16-NEXT:  .LBB97_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[8:9]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v8
-; GFX11-FAKE16-NEXT:    s_branch .LBB97_5
-; GFX11-FAKE16-NEXT:  .LBB97_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    s_branch .LBB97_2
-; GFX11-FAKE16-NEXT:  .LBB97_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s0 :: v_dual_mov_b32 v9, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s8 :: v_dual_mov_b32 v2, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s6 :: v_dual_mov_b32 v6, s5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v7, s3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s2
-; GFX11-FAKE16-NEXT:  .LBB97_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v9
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4i16_to_v8i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB97_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s3, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s7, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB97_4
+; GFX11-NEXT:  .LBB97_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v9
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[8:9]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v8
+; GFX11-NEXT:    s_branch .LBB97_5
+; GFX11-NEXT:  .LBB97_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    s_branch .LBB97_2
+; GFX11-NEXT:  .LBB97_4:
+; GFX11-NEXT:    v_dual_mov_b32 v8, s0 :: v_dual_mov_b32 v9, s1
+; GFX11-NEXT:    v_dual_mov_b32 v1, s8 :: v_dual_mov_b32 v2, s7
+; GFX11-NEXT:    v_dual_mov_b32 v5, s6 :: v_dual_mov_b32 v6, s5
+; GFX11-NEXT:    v_mov_b32_e32 v7, s3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s2
+; GFX11-NEXT:  .LBB97_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v8
+; GFX11-NEXT:    v_mov_b32_e32 v4, v9
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -15494,126 +14959,65 @@ define inreg <4 x half> @bitcast_v4bf16_to_v4f16_scalar(<4 x bfloat> inreg %a, i
 ; GFX9-NEXT:    v_mov_b32_e32 v1, s17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4bf16_to_v4f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB103_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB103_4
-; GFX11-TRUE16-NEXT:  .LBB103_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s2, 0, s1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v5
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v4, v8 :: v_dual_and_b32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v4, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v2
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB103_3:
-; GFX11-TRUE16-NEXT:    s_branch .LBB103_2
-; GFX11-TRUE16-NEXT:  .LBB103_4:
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4bf16_to_v4f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB103_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB103_4
-; GFX11-FAKE16-NEXT:  .LBB103_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s2, 0, s1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s0, 16
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v4, v8 :: v_dual_and_b32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v4, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v2
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB103_3:
-; GFX11-FAKE16-NEXT:    s_branch .LBB103_2
-; GFX11-FAKE16-NEXT:  .LBB103_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4bf16_to_v4f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s2, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB103_3
+; GFX11-NEXT:  ; %bb.1: ; %Flow
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
+; GFX11-NEXT:    s_cbranch_vccnz .LBB103_4
+; GFX11-NEXT:  .LBB103_2: ; %cmp.true
+; GFX11-NEXT:    s_pack_lh_b32_b16 s2, 0, s1
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
+; GFX11-NEXT:    s_lshl_b32 s1, s0, 16
+; GFX11-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s1
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v3, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v2
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_dual_cndmask_b32 v3, v4, v8 :: v_dual_and_b32 v2, 0xffff, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshl_or_b32 v1, v4, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v0, v0, 16, v2
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB103_3:
+; GFX11-NEXT:    s_branch .LBB103_2
+; GFX11-NEXT:  .LBB103_4:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -16021,99 +15425,50 @@ define inreg <8 x i8> @bitcast_v4f16_to_v8i8_scalar(<4 x half> inreg %a, i32 inr
 ; GFX9-NEXT:    v_mov_b32_e32 v4, v9
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4f16_to_v8i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB105_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB105_4
-; GFX11-TRUE16-NEXT:  .LBB105_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB105_5
-; GFX11-TRUE16-NEXT:  .LBB105_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB105_2
-; GFX11-TRUE16-NEXT:  .LBB105_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:  .LBB105_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4f16_to_v8i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB105_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB105_4
-; GFX11-FAKE16-NEXT:  .LBB105_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[8:9]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v8
-; GFX11-FAKE16-NEXT:    s_branch .LBB105_5
-; GFX11-FAKE16-NEXT:  .LBB105_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    s_branch .LBB105_2
-; GFX11-FAKE16-NEXT:  .LBB105_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s0 :: v_dual_mov_b32 v9, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, s8 :: v_dual_mov_b32 v2, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s6 :: v_dual_mov_b32 v6, s5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v7, s3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s2
-; GFX11-FAKE16-NEXT:  .LBB105_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v9
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4f16_to_v8i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB105_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s3, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s5, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s6, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s7, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB105_4
+; GFX11-NEXT:  .LBB105_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v9
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[8:9]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v8
+; GFX11-NEXT:    s_branch .LBB105_5
+; GFX11-NEXT:  .LBB105_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    s_branch .LBB105_2
+; GFX11-NEXT:  .LBB105_4:
+; GFX11-NEXT:    v_dual_mov_b32 v8, s0 :: v_dual_mov_b32 v9, s1
+; GFX11-NEXT:    v_dual_mov_b32 v1, s8 :: v_dual_mov_b32 v2, s7
+; GFX11-NEXT:    v_dual_mov_b32 v5, s6 :: v_dual_mov_b32 v6, s5
+; GFX11-NEXT:    v_mov_b32_e32 v7, s3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s2
+; GFX11-NEXT:  .LBB105_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v8
+; GFX11-NEXT:    v_mov_b32_e32 v4, v9
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -17344,179 +16699,88 @@ define inreg <8 x i8> @bitcast_v4bf16_to_v8i8_scalar(<4 x bfloat> inreg %a, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v4, s17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v4bf16_to_v8i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB109_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s0, 8
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB109_4
-; GFX11-TRUE16-NEXT:  .LBB109_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s0
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s3, s1, 16
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s0, 0, s1
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v4, v10 :: v_dual_add_nc_u32 v7, 0x7fff, v7
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v7, v9, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v5, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v4, v8, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v6, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v1, 16, v3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[10:11]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v11
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB109_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr5_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr2_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr7_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB109_2
-; GFX11-TRUE16-NEXT:  .LBB109_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v4bf16_to_v8i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s2, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s4, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB109_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s6, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s7, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s0, 8
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB109_4
-; GFX11-FAKE16-NEXT:  .LBB109_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s2, 0, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s1, 16
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v5
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, v3, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v4, v8, vcc_lo
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v2, 16, v3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v6, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v9
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[9:10]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v8
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB109_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr5
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr2
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr7
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB109_2
-; GFX11-FAKE16-NEXT:  .LBB109_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s8 :: v_dual_mov_b32 v7, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v5, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v1, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v4bf16_to_v8i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s2, 0
+; GFX11-NEXT:    s_mov_b32 s4, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB109_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b64 s[2:3], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s6, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s8, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s7, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s5, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s3, s0, 8
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX11-NEXT:    s_cbranch_vccnz .LBB109_4
+; GFX11-NEXT:  .LBB109_2: ; %cmp.true
+; GFX11-NEXT:    s_pack_lh_b32_b16 s2, 0, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s1, 16
+; GFX11-NEXT:    s_pack_lh_b32_b16 s1, 0, s1
+; GFX11-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
+; GFX11-NEXT:    v_bfe_u32 v4, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v7, 0x400000, v0
+; GFX11-NEXT:    v_bfe_u32 v3, v0, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v6, v2, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v8, v5, 16, 1
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, v4, v1
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v8, v8, v5
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v4, v9 :: v_dual_add_nc_u32 v6, v6, v2
+; GFX11-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
+; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v6, v10 :: v_dual_add_nc_u32 v3, v3, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v4, v8, vcc_lo
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v9, v2, 16, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshl_or_b32 v10, v6, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v9
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[9:10]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v10
+; GFX11-NEXT:    v_mov_b32_e32 v4, v8
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB109_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    ; implicit-def: $sgpr5
+; GFX11-NEXT:    ; implicit-def: $sgpr2
+; GFX11-NEXT:    ; implicit-def: $sgpr7
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB109_2
+; GFX11-NEXT:  .LBB109_4:
+; GFX11-NEXT:    v_dual_mov_b32 v6, s8 :: v_dual_mov_b32 v7, s6
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v5, s7
+; GFX11-NEXT:    v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v1, s3
+; GFX11-NEXT:    v_dual_mov_b32 v3, s2 :: v_dual_mov_b32 v4, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.704bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.704bit.ll
index 2a96722ccce0b..9f5c9c4c509ed 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.704bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.704bit.ll
@@ -3634,290 +3634,147 @@ define inreg <44 x i16> @bitcast_v22i32_to_v44i16_scalar(<22 x i32> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr14
 ; GFX9-NEXT:    s_branch .LBB13_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v22i32_to_v44i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-TRUE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB13_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s26, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s27, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s26, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s27, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s7 :: v_dual_mov_b32 v19, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB13_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB13_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v22i32_to_v44i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v3
-; GFX11-FAKE16-NEXT:    s_mov_b32 s62, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-FAKE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB13_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s28, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s29, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s6
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB13_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    s_branch .LBB13_2
+; GFX11-LABEL: bitcast_v22i32_to_v44i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v3
+; GFX11-NEXT:    s_mov_b32 s62, 0
+; GFX11-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB13_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s8, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
+; GFX11-NEXT:    s_cbranch_vccnz .LBB13_3
+; GFX11-NEXT:  .LBB13_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s8, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:  .LBB13_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s28, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s29, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s8
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
+; GFX11-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s6
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB13_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    s_branch .LBB13_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -5475,156 +5332,96 @@ define inreg <22 x i32> @bitcast_v44i16_to_v22i32_scalar(<44 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v44i16_to_v22i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v33, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v1 :: v_dual_mov_b32 v35, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v3
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB15_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s73
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v12, s75
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s0
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB15_3
 ; GFX11-TRUE16-NEXT:  .LBB15_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v35.h
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s2, s43
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
@@ -7375,290 +7172,147 @@ define inreg <44 x half> @bitcast_v22i32_to_v44f16_scalar(<22 x i32> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr14
 ; GFX9-NEXT:    s_branch .LBB17_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v22i32_to_v44f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-TRUE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB17_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s26, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s27, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s26, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s27, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s7 :: v_dual_mov_b32 v19, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB17_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB17_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v22i32_to_v44f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v3
-; GFX11-FAKE16-NEXT:    s_mov_b32 s62, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-FAKE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB17_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s28, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s29, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s6
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB17_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    s_branch .LBB17_2
+; GFX11-LABEL: bitcast_v22i32_to_v44f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v3
+; GFX11-NEXT:    s_mov_b32 s62, 0
+; GFX11-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB17_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s8, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
+; GFX11-NEXT:    s_cbranch_vccnz .LBB17_3
+; GFX11-NEXT:  .LBB17_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s8, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:  .LBB17_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s28, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s29, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s8
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
+; GFX11-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s6
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB17_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    s_branch .LBB17_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -9487,156 +9141,96 @@ define inreg <22 x i32> @bitcast_v44f16_to_v22i32_scalar(<44 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v44f16_to_v22i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v33, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v1 :: v_dual_mov_b32 v35, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v3
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB19_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s73
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v12, s75
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s0
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB19_3
 ; GFX11-TRUE16-NEXT:  .LBB19_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v35.h
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s2, s43
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
@@ -12509,305 +12103,155 @@ define inreg <44 x i16> @bitcast_v22f32_to_v44i16_scalar(<22 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr30
 ; GFX9-NEXT:    s_branch .LBB29_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v22f32_to_v44i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v21, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v19, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v5, s18
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s17 :: v_dual_mov_b32 v11, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v7, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s24 :: v_dual_mov_b32 v15, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v13, s28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s27 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-TRUE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-TRUE16-NEXT:  .LBB29_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v48, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v36, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v50, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v49, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v37, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v31, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v33, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v32, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v22
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB29_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB29_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v22f32_to_v44i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v21, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v19, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v5, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s17 :: v_dual_mov_b32 v11, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v7, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s24 :: v_dual_mov_b32 v14, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s26 :: v_dual_mov_b32 v12, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-FAKE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v12, 1.0, v12 :: v_dual_add_f32 v13, 1.0, v13
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-FAKE16-NEXT:  .LBB29_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v48, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v36, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v50, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v49, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v37, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v32, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v33, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, v22
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB29_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    s_branch .LBB29_2
+; GFX11-LABEL: bitcast_v22f32_to_v44i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
+; GFX11-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v21, s1
+; GFX11-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v19, s3
+; GFX11-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v5, s18
+; GFX11-NEXT:    v_dual_mov_b32 v6, s17 :: v_dual_mov_b32 v11, s19
+; GFX11-NEXT:    v_dual_mov_b32 v10, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v7, s23
+; GFX11-NEXT:    v_dual_mov_b32 v15, s24 :: v_dual_mov_b32 v14, s25
+; GFX11-NEXT:    v_dual_mov_b32 v13, s26 :: v_dual_mov_b32 v12, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB29_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB29_3
+; GFX11-NEXT:  .LBB29_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
+; GFX11-NEXT:    v_dual_add_f32 v12, 1.0, v12 :: v_dual_add_f32 v13, 1.0, v13
+; GFX11-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v6, 1.0, v6
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
+; GFX11-NEXT:  .LBB29_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v23, v23, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v25, v25, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v4, v4, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v7, v48, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v11, v36, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v5, v50, 16, v6
+; GFX11-NEXT:    v_lshl_or_b32 v6, v49, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshl_or_b32 v8, v39, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v37, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_lshl_or_b32 v12, v35, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v15, v32, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v25
+; GFX11-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
+; GFX11-NEXT:    v_mov_b32_e32 v3, v23
+; GFX11-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v9, v38, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v13, v34, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v14, v33, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v16, v31, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
+; GFX11-NEXT:    v_mov_b32_e32 v0, v24
+; GFX11-NEXT:    v_mov_b32_e32 v2, v22
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB29_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr24
+; GFX11-NEXT:    ; implicit-def: $vgpr25
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr23
+; GFX11-NEXT:    ; implicit-def: $vgpr4
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    s_branch .LBB29_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -14365,156 +13809,96 @@ define inreg <22 x float> @bitcast_v44i16_to_v22f32_scalar(<44 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v44i16_to_v22f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v33, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v1 :: v_dual_mov_b32 v35, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v3
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB31_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s73
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v12, s75
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s0
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB31_3
 ; GFX11-TRUE16-NEXT:  .LBB31_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v35.h
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s2, s43
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
@@ -16250,305 +15634,155 @@ define inreg <44 x half> @bitcast_v22f32_to_v44f16_scalar(<22 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr30
 ; GFX9-NEXT:    s_branch .LBB33_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v22f32_to_v44f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v21, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v19, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v5, s18
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s17 :: v_dual_mov_b32 v11, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v7, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s24 :: v_dual_mov_b32 v15, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v13, s28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s27 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-TRUE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-TRUE16-NEXT:  .LBB33_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v48, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v36, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v50, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v49, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v37, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v31, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v33, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v32, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v22
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB33_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB33_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v22f32_to_v44f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v21, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v19, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v5, s18
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s17 :: v_dual_mov_b32 v11, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v7, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s24 :: v_dual_mov_b32 v14, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s26 :: v_dual_mov_b32 v12, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-FAKE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v12, 1.0, v12 :: v_dual_add_f32 v13, 1.0, v13
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-FAKE16-NEXT:  .LBB33_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v48, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v36, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v50, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v49, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v37, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v32, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v33, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, v22
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB33_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr23
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    s_branch .LBB33_2
+; GFX11-LABEL: bitcast_v22f32_to_v44f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
+; GFX11-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v21, s1
+; GFX11-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v19, s3
+; GFX11-NEXT:    v_dual_mov_b32 v18, s16 :: v_dual_mov_b32 v5, s18
+; GFX11-NEXT:    v_dual_mov_b32 v6, s17 :: v_dual_mov_b32 v11, s19
+; GFX11-NEXT:    v_dual_mov_b32 v10, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v7, s23
+; GFX11-NEXT:    v_dual_mov_b32 v15, s24 :: v_dual_mov_b32 v14, s25
+; GFX11-NEXT:    v_dual_mov_b32 v13, s26 :: v_dual_mov_b32 v12, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB33_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB33_3
+; GFX11-NEXT:  .LBB33_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
+; GFX11-NEXT:    v_dual_add_f32 v12, 1.0, v12 :: v_dual_add_f32 v13, 1.0, v13
+; GFX11-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v8, 1.0, v8
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v6, 1.0, v6
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
+; GFX11-NEXT:  .LBB33_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v23, v23, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v25, v25, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v4, v4, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v7, v48, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v11, v36, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v5, v50, 16, v6
+; GFX11-NEXT:    v_lshl_or_b32 v6, v49, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_lshl_or_b32 v8, v39, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v37, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_lshl_or_b32 v12, v35, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v15, v32, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
+; GFX11-NEXT:    v_mov_b32_e32 v1, v25
+; GFX11-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
+; GFX11-NEXT:    v_mov_b32_e32 v3, v23
+; GFX11-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v9, v38, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v13, v34, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v14, v33, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v16, v31, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
+; GFX11-NEXT:    v_mov_b32_e32 v0, v24
+; GFX11-NEXT:    v_mov_b32_e32 v2, v22
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB33_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr24
+; GFX11-NEXT:    ; implicit-def: $vgpr25
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr23
+; GFX11-NEXT:    ; implicit-def: $vgpr4
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    s_branch .LBB33_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -18377,156 +17611,96 @@ define inreg <22 x float> @bitcast_v44f16_to_v22f32_scalar(<44 x half> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v44f16_to_v22f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v33, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v1 :: v_dual_mov_b32 v35, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v3
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB35_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s73
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v12, s75
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s0
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB35_3
 ; GFX11-TRUE16-NEXT:  .LBB35_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v35.h
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s2, s43
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
@@ -20704,290 +19878,147 @@ define inreg <44 x i16> @bitcast_v11i64_to_v44i16_scalar(<11 x i64> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr14
 ; GFX9-NEXT:    s_branch .LBB41_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v11i64_to_v44i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-TRUE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB41_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s26, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s27, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s26, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s27, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s7 :: v_dual_mov_b32 v19, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB41_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB41_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v11i64_to_v44i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v3
-; GFX11-FAKE16-NEXT:    s_mov_b32 s62, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-FAKE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s6, s6, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB41_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s28, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s29, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s6
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB41_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    s_branch .LBB41_2
+; GFX11-LABEL: bitcast_v11i64_to_v44i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v3
+; GFX11-NEXT:    s_mov_b32 s62, 0
+; GFX11-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB41_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s8, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
+; GFX11-NEXT:    s_cbranch_vccnz .LBB41_3
+; GFX11-NEXT:  .LBB41_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s7, s7, 3
+; GFX11-NEXT:    s_addc_u32 s6, s6, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s8, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:  .LBB41_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s28, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s29, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s8
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
+; GFX11-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s6
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB41_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    s_branch .LBB41_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -22545,156 +21576,96 @@ define inreg <11 x i64> @bitcast_v44i16_to_v11i64_scalar(<44 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v44i16_to_v11i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v33, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v1 :: v_dual_mov_b32 v35, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v3
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB43_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s73
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v12, s75
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s0
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB43_3
 ; GFX11-TRUE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v35.h
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s2, s43
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
@@ -24457,290 +23428,147 @@ define inreg <44 x half> @bitcast_v11i64_to_v44f16_scalar(<11 x i64> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr14
 ; GFX9-NEXT:    s_branch .LBB45_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v11i64_to_v44f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-TRUE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB45_3: ; %end
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s26, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s27, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s26, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s27, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s7 :: v_dual_mov_b32 v19, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB45_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB45_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v11i64_to_v44f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v3
-; GFX11-FAKE16-NEXT:    s_mov_b32 s62, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-FAKE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s6, s6, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB45_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s28, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s29, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s9
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s6
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB45_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    s_branch .LBB45_2
+; GFX11-LABEL: bitcast_v11i64_to_v44f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v3
+; GFX11-NEXT:    s_mov_b32 s62, 0
+; GFX11-NEXT:    s_and_b32 s8, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB45_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s8, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s62
+; GFX11-NEXT:    s_cbranch_vccnz .LBB45_3
+; GFX11-NEXT:  .LBB45_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s7, s7, 3
+; GFX11-NEXT:    s_addc_u32 s6, s6, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s8, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s0, 16
+; GFX11-NEXT:  .LBB45_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s27, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s28, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s29, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s10
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s9
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s8
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s14
+; GFX11-NEXT:    v_dual_mov_b32 v16, s13 :: v_dual_mov_b32 v17, s12
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s6
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB45_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    s_branch .LBB45_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -26569,156 +25397,96 @@ define inreg <11 x i64> @bitcast_v44f16_to_v11i64_scalar(<44 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v44f16_to_v11i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v33, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v1 :: v_dual_mov_b32 v35, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v3
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB47_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s73
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v12, s75
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s0
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB47_3
 ; GFX11-TRUE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v35.h
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s2, s43
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
@@ -28033,303 +26801,154 @@ define inreg <44 x i16> @bitcast_v11f64_to_v44i16_scalar(<11 x double> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr30
 ; GFX9-NEXT:    s_branch .LBB49_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v11f64_to_v44i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v23, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v21, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s22 :: v_dual_mov_b32 v8, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v15, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v13, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-TRUE16-NEXT:  .LBB49_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v48, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v39, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v37, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v36, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v11, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v33, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v32, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v31, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v22
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v11f64_to_v44i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v23, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v21, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s22 :: v_dual_mov_b32 v8, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v15, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v13, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-FAKE16-NEXT:  .LBB49_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v48, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v39, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v37, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v36, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v11, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v33, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v32, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, v22
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
+; GFX11-LABEL: bitcast_v11f64_to_v44i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
+; GFX11-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v23, s1
+; GFX11-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v21, s3
+; GFX11-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
+; GFX11-NEXT:    v_dual_mov_b32 v7, s22 :: v_dual_mov_b32 v8, s23
+; GFX11-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v15, s25
+; GFX11-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v13, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_3
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
+; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
+; GFX11-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
+; GFX11-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
+; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
+; GFX11-NEXT:  .LBB49_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v23, v50, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
+; GFX11-NEXT:    v_lshl_or_b32 v5, v49, 16, v6
+; GFX11-NEXT:    v_lshl_or_b32 v6, v48, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v7, v39, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, v23
+; GFX11-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v8, v38, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v9, v37, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v36, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v11, v11, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v12, v35, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v13, v34, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v14, v33, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v15, v32, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v16, v31, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
+; GFX11-NEXT:    v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
+; GFX11-NEXT:    v_mov_b32_e32 v2, v22
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr24
+; GFX11-NEXT:    ; implicit-def: $vgpr25
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr4
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr11
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    s_branch .LBB49_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -29887,156 +28506,96 @@ define inreg <11 x double> @bitcast_v44i16_to_v11f64_scalar(<44 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v44i16_to_v11f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v33, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v1 :: v_dual_mov_b32 v35, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v3
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB51_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s73
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v12, s75
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s0
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB51_3
 ; GFX11-TRUE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v35.h
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s2, s43
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
@@ -31697,303 +30256,154 @@ define inreg <44 x half> @bitcast_v11f64_to_v44f16_scalar(<11 x double> inreg %a
 ; GFX9-NEXT:    ; implicit-def: $vgpr30
 ; GFX9-NEXT:    s_branch .LBB53_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v11f64_to_v44f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v23, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v21, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s22 :: v_dual_mov_b32 v8, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v15, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v13, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-TRUE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-TRUE16-NEXT:  .LBB53_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v48, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v39, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v37, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v36, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v11, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v33, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v32, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v31, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v22
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB53_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB53_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v11f64_to_v44f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v23, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v21, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s22 :: v_dual_mov_b32 v8, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v15, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v13, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-FAKE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
-; GFX11-FAKE16-NEXT:  .LBB53_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v48, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v39, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v37, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v36, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v11, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v33, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v32, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v31, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, v22
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB53_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr24
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    s_branch .LBB53_2
+; GFX11-LABEL: bitcast_v11f64_to_v44f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
+; GFX11-NEXT:    v_dual_mov_b32 v22, s0 :: v_dual_mov_b32 v23, s1
+; GFX11-NEXT:    v_dual_mov_b32 v20, s2 :: v_dual_mov_b32 v21, s3
+; GFX11-NEXT:    v_dual_mov_b32 v5, s16 :: v_dual_mov_b32 v6, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
+; GFX11-NEXT:    v_dual_mov_b32 v7, s22 :: v_dual_mov_b32 v8, s23
+; GFX11-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v15, s25
+; GFX11-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v13, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB53_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB53_3
+; GFX11-NEXT:  .LBB53_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
+; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
+; GFX11-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
+; GFX11-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
+; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
+; GFX11-NEXT:  .LBB53_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v23, v50, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
+; GFX11-NEXT:    v_lshl_or_b32 v5, v49, 16, v6
+; GFX11-NEXT:    v_lshl_or_b32 v6, v48, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v24, v24, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v7, v39, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, v23
+; GFX11-NEXT:    v_lshl_or_b32 v22, v51, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v8, v38, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v9, v37, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v36, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v11, v11, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v12, v35, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v13, v34, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v14, v33, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v15, v32, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v16, v31, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v30, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v18, v29, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v19, v28, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v20, v27, 16, v2
+; GFX11-NEXT:    v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
+; GFX11-NEXT:    v_mov_b32_e32 v2, v22
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB53_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr24
+; GFX11-NEXT:    ; implicit-def: $vgpr25
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr4
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr11
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    s_branch .LBB53_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -33822,156 +32232,96 @@ define inreg <11 x double> @bitcast_v44f16_to_v11f64_scalar(<44 x half> inreg %a
 ; GFX11-TRUE16-LABEL: bitcast_v44f16_to_v11f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v3 :: v_dual_mov_b32 v33, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v1 :: v_dual_mov_b32 v35, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v3
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s26, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB55_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s73
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v12, s75
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s15 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s0
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB55_3
 ; GFX11-TRUE16-NEXT:  .LBB55_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v35.h
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s2, s43
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v36
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
@@ -35941,7 +34291,15 @@ define inreg <44 x half> @bitcast_v44i16_to_v44f16_scalar(<44 x i16> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v44i16_to_v44f16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v21.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v21.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v21.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
@@ -35951,15 +34309,15 @@ define inreg <44 x half> @bitcast_v44i16_to_v44f16_scalar(<44 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB57_3
@@ -35967,181 +34325,146 @@ define inreg <44 x half> @bitcast_v44i16_to_v44f16_scalar(<44 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB57_4
 ; GFX11-TRUE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v0.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v2.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v7, 16, v0
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s15, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v5, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v6, 16, v1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s15, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s14, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s13, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s12, s9
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s11, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s9, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s21, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s16, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s29, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s28, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s27, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s26, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s25, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s24, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s9, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s8, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v25, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v24, s1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v23, s2, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v22, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v25
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v24
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v23
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v22
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v13
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v48.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v49.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v50.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v51.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_5
 ; GFX11-TRUE16-NEXT:  .LBB57_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_2
 ; GFX11-TRUE16-NEXT:  .LBB57_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s29 :: v_dual_mov_b32 v12, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s27 :: v_dual_mov_b32 v14, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s25 :: v_dual_mov_b32 v16, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s23 :: v_dual_mov_b32 v8, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s21 :: v_dual_mov_b32 v10, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s19 :: v_dual_mov_b32 v4, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s3 :: v_dual_mov_b32 v23, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s1 :: v_dual_mov_b32 v25, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s45 :: v_dual_mov_b32 v27, s44
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s43 :: v_dual_mov_b32 v29, s42
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s41 :: v_dual_mov_b32 v31, s40
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s15 :: v_dual_mov_b32 v33, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s13 :: v_dual_mov_b32 v35, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, s10 :: v_dual_mov_b32 v37, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s8 :: v_dual_mov_b32 v39, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, s5 :: v_dual_mov_b32 v49, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, s4 :: v_dual_mov_b32 v51, s12
 ; GFX11-TRUE16-NEXT:  .LBB57_5: ; %end
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v24
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v51, 16, v25
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v50, 16, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v49, 16, v23
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v48, 16, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v38, 16, v5
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v49
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v28, 16, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v23
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v37, 16, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v36, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v35, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v33, 16, v37
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v38, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v39, 16, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v29, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v21, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v26, 16, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v12
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v2.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v3.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v36, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v30, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v27, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v34, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v31, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v30, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v29, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v28, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v27, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v26, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v2
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v22 :: v_dual_mov_b32 v3, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v37, 16, v38
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v22
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: bitcast_v44i16_to_v44f16_scalar:
@@ -37919,7 +36242,15 @@ define inreg <44 x i16> @bitcast_v44f16_to_v44i16_scalar(<44 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v44f16_to_v44i16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v4
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v21.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v21.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v21.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
@@ -37929,15 +36260,15 @@ define inreg <44 x i16> @bitcast_v44f16_to_v44i16_scalar(<44 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB59_3
@@ -37945,181 +36276,146 @@ define inreg <44 x i16> @bitcast_v44f16_to_v44i16_scalar(<44 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB59_4
 ; GFX11-TRUE16-NEXT:  .LBB59_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v0.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v2.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v7, 16, v0
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v5, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v6, 16, v1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s15, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s14, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s13, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s12, s9
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s11, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s9, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s21, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s16, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s29 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s28 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s27 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s26 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s25 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s24 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s9 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s8 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v25, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v24, 0x200, s1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v23, 0x200, s2 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v22, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v25
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v24
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v23
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v22
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v13
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v48.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v49.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v50.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v51.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_5
 ; GFX11-TRUE16-NEXT:  .LBB59_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_2
 ; GFX11-TRUE16-NEXT:  .LBB59_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s29 :: v_dual_mov_b32 v12, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s27 :: v_dual_mov_b32 v14, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s25 :: v_dual_mov_b32 v16, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s23 :: v_dual_mov_b32 v8, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s21 :: v_dual_mov_b32 v10, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s19 :: v_dual_mov_b32 v4, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s17 :: v_dual_mov_b32 v6, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s3 :: v_dual_mov_b32 v23, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s1 :: v_dual_mov_b32 v25, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s45 :: v_dual_mov_b32 v27, s44
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s43 :: v_dual_mov_b32 v29, s42
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s41 :: v_dual_mov_b32 v31, s40
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s15 :: v_dual_mov_b32 v33, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s13 :: v_dual_mov_b32 v35, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, s10 :: v_dual_mov_b32 v37, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s8 :: v_dual_mov_b32 v39, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, s5 :: v_dual_mov_b32 v49, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, s4 :: v_dual_mov_b32 v51, s12
 ; GFX11-TRUE16-NEXT:  .LBB59_5: ; %end
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v24
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v22
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v51, 16, v25
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v50, 16, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v49, 16, v23
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v48, 16, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v38, 16, v5
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v49
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v28, 16, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v23
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v37, 16, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v36, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v35, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v33, 16, v37
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v38, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v39, 16, v48
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v29, 16, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v19, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v20, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v21, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v26, 16, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v12
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v2.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v3.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v36, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v30, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v32, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v27, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v34, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v31, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v30, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v29, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v28, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v27, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v26, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v2
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v22 :: v_dual_mov_b32 v3, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v37, 16, v38
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v22
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: bitcast_v44f16_to_v44i16_scalar:
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.768bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.768bit.ll
index a2bd1d30cc634..152a48bec2636 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.768bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.768bit.ll
@@ -3858,315 +3858,160 @@ define inreg <48 x i16> @bitcast_v24i32_to_v48i16_scalar(<24 x i32> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr40
 ; GFX9-NEXT:    s_branch .LBB13_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v24i32_to_v48i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-TRUE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB13_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s9 :: v_dual_mov_b32 v19, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s6 :: v_dual_mov_b32 v23, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB13_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB13_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v24i32_to_v48i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-FAKE16-NEXT:    s_mov_b32 s74, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-FAKE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB13_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB13_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    s_branch .LBB13_2
+; GFX11-LABEL: bitcast_v24i32_to_v48i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v5
+; GFX11-NEXT:    s_mov_b32 s74, 0
+; GFX11-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB13_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s10, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
+; GFX11-NEXT:    s_cbranch_vccnz .LBB13_3
+; GFX11-NEXT:  .LBB13_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s8, s8, 3
+; GFX11-NEXT:    s_add_i32 s9, s9, 3
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s10, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s0, 16
+; GFX11-NEXT:  .LBB13_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s10
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB13_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    s_branch .LBB13_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -5950,172 +5795,107 @@ define inreg <24 x i32> @bitcast_v48i16_to_v24i32_scalar(<48 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v24i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v5 :: v_dual_mov_b32 v33, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v5
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s26, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB15_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s57
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB15_3
 ; GFX11-TRUE16-NEXT:  .LBB15_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v28
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
@@ -8045,315 +7825,160 @@ define inreg <48 x half> @bitcast_v24i32_to_v48f16_scalar(<24 x i32> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr40
 ; GFX9-NEXT:    s_branch .LBB17_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v24i32_to_v48f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-TRUE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB17_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s9 :: v_dual_mov_b32 v19, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s6 :: v_dual_mov_b32 v23, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB17_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB17_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v24i32_to_v48f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-FAKE16-NEXT:    s_mov_b32 s74, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-FAKE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB17_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB17_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    s_branch .LBB17_2
+; GFX11-LABEL: bitcast_v24i32_to_v48f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v5
+; GFX11-NEXT:    s_mov_b32 s74, 0
+; GFX11-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB17_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s10, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
+; GFX11-NEXT:    s_cbranch_vccnz .LBB17_3
+; GFX11-NEXT:  .LBB17_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s8, s8, 3
+; GFX11-NEXT:    s_add_i32 s9, s9, 3
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s10, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s0, 16
+; GFX11-NEXT:  .LBB17_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s10
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB17_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    s_branch .LBB17_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -10409,172 +10034,107 @@ define inreg <24 x i32> @bitcast_v48f16_to_v24i32_scalar(<48 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v24i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v5 :: v_dual_mov_b32 v33, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v5
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s26, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB19_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s57
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB19_3
 ; GFX11-TRUE16-NEXT:  .LBB19_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v28
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
@@ -13634,327 +13194,166 @@ define inreg <48 x i16> @bitcast_v24f32_to_v48i16_scalar(<24 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr34
 ; GFX9-NEXT:    s_branch .LBB29_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v24f32_to_v48i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v23, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v21, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v19, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v7, s20
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s19 :: v_dual_mov_b32 v13, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v9, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v16, s28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v14, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-TRUE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-TRUE16-NEXT:  .LBB29_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v51, 16, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v49, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v39, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v36, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v53, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v52, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v48, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v38, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v37, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB29_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB29_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v24f32_to_v48i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v23, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v21, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v19, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v7, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s19 :: v_dual_mov_b32 v13, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v9, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v14, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-FAKE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-FAKE16-NEXT:  .LBB29_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v51, 16, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v49, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v39, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v38, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v52, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v48, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v53, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v37, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v36, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB29_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    s_branch .LBB29_2
+; GFX11-LABEL: bitcast_v24f32_to_v48i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX11-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v23, s1
+; GFX11-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v21, s3
+; GFX11-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v19, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v7, s20
+; GFX11-NEXT:    v_dual_mov_b32 v8, s19 :: v_dual_mov_b32 v13, s21
+; GFX11-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v9, s25
+; GFX11-NEXT:    v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v14, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB29_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB29_3
+; GFX11-NEXT:  .LBB29_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
+; GFX11-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
+; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v8, 1.0, v8
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
+; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
+; GFX11-NEXT:  .LBB29_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_lshl_or_b32 v6, v6, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v51, 16, v12
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v12, v49, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v27, v27, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v9
+; GFX11-NEXT:    v_lshl_or_b32 v14, v39, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v15, v38, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_lshl_or_b32 v9, v52, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v13, v48, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v4
+; GFX11-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v25
+; GFX11-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v7, v54, 16, v8
+; GFX11-NEXT:    v_lshl_or_b32 v8, v53, 16, v21
+; GFX11-NEXT:    v_lshl_or_b32 v11, v50, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v16, v37, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v36, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v24
+; GFX11-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
+; GFX11-NEXT:    v_mov_b32_e32 v4, v28
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB29_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr25
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr6
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    s_branch .LBB29_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -15738,172 +15137,107 @@ define inreg <24 x float> @bitcast_v48i16_to_v24f32_scalar(<48 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v24f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v5 :: v_dual_mov_b32 v33, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v5
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s26, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB31_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s57
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB31_3
 ; GFX11-TRUE16-NEXT:  .LBB31_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v28
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
@@ -17817,327 +17151,166 @@ define inreg <48 x half> @bitcast_v24f32_to_v48f16_scalar(<24 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr34
 ; GFX9-NEXT:    s_branch .LBB33_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v24f32_to_v48f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v23, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v21, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v19, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v7, s20
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s19 :: v_dual_mov_b32 v13, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v9, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v16, s28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v14, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-TRUE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-TRUE16-NEXT:  .LBB33_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v51, 16, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v49, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v39, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v36, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v53, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v52, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v48, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v38, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v37, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB33_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB33_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v24f32_to_v48f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v23, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v21, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v19, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v7, s20
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s19 :: v_dual_mov_b32 v13, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v9, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v14, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-FAKE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-FAKE16-NEXT:  .LBB33_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v51, 16, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v49, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v39, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v38, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v52, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v48, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v53, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v37, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v36, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB33_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr25
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    s_branch .LBB33_2
+; GFX11-LABEL: bitcast_v24f32_to_v48f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX11-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v23, s1
+; GFX11-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v21, s3
+; GFX11-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v19, s17
+; GFX11-NEXT:    v_dual_mov_b32 v18, s18 :: v_dual_mov_b32 v7, s20
+; GFX11-NEXT:    v_dual_mov_b32 v8, s19 :: v_dual_mov_b32 v13, s21
+; GFX11-NEXT:    v_dual_mov_b32 v12, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v10, s24 :: v_dual_mov_b32 v9, s25
+; GFX11-NEXT:    v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v14, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB33_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB33_3
+; GFX11-NEXT:  .LBB33_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
+; GFX11-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v15, 1.0, v15
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v10, 1.0, v10
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
+; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v8, 1.0, v8
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
+; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
+; GFX11-NEXT:  .LBB33_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_lshl_or_b32 v6, v6, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v51, 16, v12
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v25, v25, 16, v23
+; GFX11-NEXT:    v_lshl_or_b32 v12, v49, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v27, v27, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v9
+; GFX11-NEXT:    v_lshl_or_b32 v14, v39, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v15, v38, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_lshl_or_b32 v9, v52, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v13, v48, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v4
+; GFX11-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v25
+; GFX11-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v7, v54, 16, v8
+; GFX11-NEXT:    v_lshl_or_b32 v8, v53, 16, v21
+; GFX11-NEXT:    v_lshl_or_b32 v11, v50, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v16, v37, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v36, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v24
+; GFX11-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
+; GFX11-NEXT:    v_mov_b32_e32 v4, v28
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB33_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr25
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr6
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    s_branch .LBB33_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -20193,172 +19366,107 @@ define inreg <24 x float> @bitcast_v48f16_to_v24f32_scalar(<48 x half> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v24f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v5 :: v_dual_mov_b32 v33, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v5
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s26, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB35_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s57
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB35_3
 ; GFX11-TRUE16-NEXT:  .LBB35_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v28
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
@@ -22693,315 +21801,160 @@ define inreg <48 x i16> @bitcast_v12i64_to_v48i16_scalar(<12 x i64> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr40
 ; GFX9-NEXT:    s_branch .LBB41_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v12i64_to_v48i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-TRUE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s8, s8, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB41_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s9 :: v_dual_mov_b32 v19, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s6 :: v_dual_mov_b32 v23, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB41_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB41_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v12i64_to_v48i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-FAKE16-NEXT:    s_mov_b32 s74, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-FAKE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s8, s8, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s7, s7, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB41_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB41_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    s_branch .LBB41_2
+; GFX11-LABEL: bitcast_v12i64_to_v48i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v5
+; GFX11-NEXT:    s_mov_b32 s74, 0
+; GFX11-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB41_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s10, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
+; GFX11-NEXT:    s_cbranch_vccnz .LBB41_3
+; GFX11-NEXT:  .LBB41_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s9, s9, 3
+; GFX11-NEXT:    s_addc_u32 s8, s8, 0
+; GFX11-NEXT:    s_add_u32 s6, s6, 3
+; GFX11-NEXT:    s_addc_u32 s7, s7, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s10, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s0, 16
+; GFX11-NEXT:  .LBB41_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s10
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB41_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    s_branch .LBB41_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -24785,172 +23738,107 @@ define inreg <12 x i64> @bitcast_v48i16_to_v12i64_scalar(<48 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v12i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v5 :: v_dual_mov_b32 v33, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v5
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s26, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB43_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s57
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB43_3
 ; GFX11-TRUE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v28
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
@@ -26892,315 +25780,160 @@ define inreg <48 x half> @bitcast_v12i64_to_v48f16_scalar(<12 x i64> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr40
 ; GFX9-NEXT:    s_branch .LBB45_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v12i64_to_v48f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-TRUE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s8, s8, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB45_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s9 :: v_dual_mov_b32 v19, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s7 :: v_dual_mov_b32 v21, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s6 :: v_dual_mov_b32 v23, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB45_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB45_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v12i64_to_v48f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-FAKE16-NEXT:    s_mov_b32 s74, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-FAKE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s8, s8, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s7, s7, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB45_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s12
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s11
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB45_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    s_branch .LBB45_2
+; GFX11-LABEL: bitcast_v12i64_to_v48f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v5
+; GFX11-NEXT:    s_mov_b32 s74, 0
+; GFX11-NEXT:    s_and_b32 s10, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB45_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s10, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s74
+; GFX11-NEXT:    s_cbranch_vccnz .LBB45_3
+; GFX11-NEXT:  .LBB45_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s9, s9, 3
+; GFX11-NEXT:    s_addc_u32 s8, s8, 0
+; GFX11-NEXT:    s_add_u32 s6, s6, 3
+; GFX11-NEXT:    s_addc_u32 s7, s7, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s10, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s0, 16
+; GFX11-NEXT:  .LBB45_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s12
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s11
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s10
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB45_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    s_branch .LBB45_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -29256,172 +27989,107 @@ define inreg <12 x i64> @bitcast_v48f16_to_v12i64_scalar(<48 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v12i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v5 :: v_dual_mov_b32 v33, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v5
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s26, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB47_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s57
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB47_3
 ; GFX11-TRUE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v28
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
@@ -30849,327 +29517,166 @@ define inreg <48 x i16> @bitcast_v12f64_to_v48i16_scalar(<12 x double> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr34
 ; GFX9-NEXT:    s_branch .LBB49_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v12f64_to_v48i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v25, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v23, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v21, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v19, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s24 :: v_dual_mov_b32 v10, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s26 :: v_dual_mov_b32 v17, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s28 :: v_dual_mov_b32 v15, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-TRUE16-NEXT:  .LBB49_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v52, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v51, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v39, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v37, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v49, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v48, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v38, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v36, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v12f64_to_v48i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v25, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v23, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v21, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v19, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s24 :: v_dual_mov_b32 v10, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-FAKE16-NEXT:  .LBB49_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v52, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v51, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v49, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v48, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v39, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v38, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v37, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v36, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
+; GFX11-LABEL: bitcast_v12f64_to_v48i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX11-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v25, s1
+; GFX11-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v23, s3
+; GFX11-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v21, s17
+; GFX11-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
+; GFX11-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v19, s21
+; GFX11-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
+; GFX11-NEXT:    v_dual_mov_b32 v9, s24 :: v_dual_mov_b32 v10, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_3
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
+; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
+; GFX11-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
+; GFX11-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
+; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
+; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
+; GFX11-NEXT:  .LBB49_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v6, v6, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v7, v53, 16, v8
+; GFX11-NEXT:    v_lshl_or_b32 v8, v52, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v25, v54, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v9
+; GFX11-NEXT:    v_lshl_or_b32 v9, v51, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v4
+; GFX11-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v10
+; GFX11-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v13, v13, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v25
+; GFX11-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v10, v50, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v11, v49, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v48, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v14, v39, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v15, v38, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v37, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v36, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v24
+; GFX11-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
+; GFX11-NEXT:    v_mov_b32_e32 v4, v28
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr6
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr13
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    s_branch .LBB49_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -32953,172 +31460,107 @@ define inreg <12 x double> @bitcast_v48i16_to_v12f64_scalar(<48 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v12f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v5 :: v_dual_mov_b32 v33, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v5
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s26, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB51_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s57
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB51_3
 ; GFX11-TRUE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s4, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v28
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s14, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s17, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
@@ -34950,327 +33392,166 @@ define inreg <48 x half> @bitcast_v12f64_to_v48f16_scalar(<12 x double> inreg %a
 ; GFX9-NEXT:    ; implicit-def: $vgpr34
 ; GFX9-NEXT:    s_branch .LBB53_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v12f64_to_v48f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v25, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v23, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v21, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v19, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s24 :: v_dual_mov_b32 v10, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s26 :: v_dual_mov_b32 v17, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s28 :: v_dual_mov_b32 v15, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-TRUE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-TRUE16-NEXT:  .LBB53_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v52, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v51, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v39, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v37, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v49, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v48, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v38, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v36, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB53_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB53_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v12f64_to_v48f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v25, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v23, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v21, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v19, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s24 :: v_dual_mov_b32 v10, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-FAKE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
-; GFX11-FAKE16-NEXT:  .LBB53_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v52, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v51, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v49, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v48, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v39, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v38, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v37, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v36, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB53_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr26
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr6
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    s_branch .LBB53_2
+; GFX11-LABEL: bitcast_v12f64_to_v48f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX11-NEXT:    v_dual_mov_b32 v24, s0 :: v_dual_mov_b32 v25, s1
+; GFX11-NEXT:    v_dual_mov_b32 v22, s2 :: v_dual_mov_b32 v23, s3
+; GFX11-NEXT:    v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v21, s17
+; GFX11-NEXT:    v_dual_mov_b32 v7, s18 :: v_dual_mov_b32 v8, s19
+; GFX11-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v19, s21
+; GFX11-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
+; GFX11-NEXT:    v_dual_mov_b32 v9, s24 :: v_dual_mov_b32 v10, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB53_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB53_3
+; GFX11-NEXT:  .LBB53_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
+; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
+; GFX11-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
+; GFX11-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
+; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
+; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v24
+; GFX11-NEXT:  .LBB53_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v7, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v6, v6, 16, v7
+; GFX11-NEXT:    v_lshl_or_b32 v7, v53, 16, v8
+; GFX11-NEXT:    v_lshl_or_b32 v8, v52, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v25, v54, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v9
+; GFX11-NEXT:    v_lshl_or_b32 v9, v51, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v35, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v4
+; GFX11-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v28, v28, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v10
+; GFX11-NEXT:    v_lshl_or_b32 v26, v26, 16, v22
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v13, v13, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_lshl_or_b32 v20, v33, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v25
+; GFX11-NEXT:    v_lshl_or_b32 v24, v55, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v10, v50, 16, v11
+; GFX11-NEXT:    v_lshl_or_b32 v11, v49, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v48, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v14, v39, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v15, v38, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v37, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v36, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v19, v34, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v21, v32, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v22, v31, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v23, v30, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v24
+; GFX11-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
+; GFX11-NEXT:    v_mov_b32_e32 v4, v28
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB53_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr26
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr6
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr13
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    s_branch .LBB53_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -37326,172 +35607,107 @@ define inreg <12 x double> @bitcast_v48f16_to_v12f64_scalar(<48 x half> inreg %a
 ; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v12f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v5 :: v_dual_mov_b32 v33, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v1 :: v_dual_mov_b32 v37, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v5
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s0, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s16, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s17, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s18, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s19, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s20, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s21, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s22, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s23, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s24, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s25, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s26, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s27, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s28, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB55_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s57
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s5 :: v_dual_mov_b32 v5, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v7, s8
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s9 :: v_dual_mov_b32 v9, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v11, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s13 :: v_dual_mov_b32 v13, s14
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s16 :: v_dual_mov_b32 v17, s17
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB55_3
 ; GFX11-TRUE16-NEXT:  .LBB55_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v37, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v33, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v32, 16, v38
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s4 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v28
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s14 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s17 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
@@ -39716,7 +37932,19 @@ define inreg <48 x half> @bitcast_v48i16_to_v48f16_scalar(<48 x i16> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v48i16_to_v48f16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v23.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v23.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v23.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v23.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v23.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
@@ -39724,17 +37952,17 @@ define inreg <48 x half> @bitcast_v48i16_to_v48f16_scalar(<48 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s25, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s24, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB57_3
@@ -39742,198 +37970,156 @@ define inreg <48 x half> @bitcast_v48i16_to_v48f16_scalar(<48 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB57_4
 ; GFX11-TRUE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v5.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v4.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v2.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v7, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s15, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v8, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v6, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v7, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v9, 16, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s15, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s14, s11
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s13, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s11, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s16, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s10, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s7, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s29, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s28, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s27, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s26, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s25, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s24, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v28, s8, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v27, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v26, s1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v25, s2, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v24, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v23, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v29, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v26
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v25
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v24
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v28
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v10
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v50.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v51.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v52.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v53.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v54.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v55.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v5
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_5
 ; GFX11-TRUE16-NEXT:  .LBB57_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_2
 ; GFX11-TRUE16-NEXT:  .LBB57_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s29 :: v_dual_mov_b32 v15, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s27 :: v_dual_mov_b32 v17, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s25 :: v_dual_mov_b32 v10, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v12, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s21 :: v_dual_mov_b32 v6, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v8, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s17 :: v_dual_mov_b32 v29, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s3 :: v_dual_mov_b32 v25, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s1 :: v_dual_mov_b32 v27, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s45 :: v_dual_mov_b32 v31, s44
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s43 :: v_dual_mov_b32 v33, s42
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s41 :: v_dual_mov_b32 v35, s40
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, s15 :: v_dual_mov_b32 v37, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s12 :: v_dual_mov_b32 v39, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, s10 :: v_dual_mov_b32 v49, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, s8 :: v_dual_mov_b32 v51, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s5 :: v_dual_mov_b32 v53, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, s4 :: v_dual_mov_b32 v55, s14
 ; GFX11-TRUE16-NEXT:  .LBB57_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v33, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v38, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v49, 16, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v24
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v20, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v39, 16, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v27
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v52, 16, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v25
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v26
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v30, 16, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v28
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v3
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v48, 16, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v37, 16, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v32, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v53, 16, v64
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v49, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v48, 16, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v50, 16, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v38, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v37, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v36, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v33, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v32, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v31, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v30, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v4
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v26
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v28 :: v_dual_mov_b32 v5, v29
@@ -41911,7 +40097,19 @@ define inreg <48 x i16> @bitcast_v48f16_to_v48i16_scalar(<48 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v48f16_to_v48i16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v6
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v23.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v23.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v23.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v23.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v23.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
@@ -41919,17 +40117,17 @@ define inreg <48 x i16> @bitcast_v48f16_to_v48i16_scalar(<48 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s25, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s24, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB59_3
@@ -41937,198 +40135,156 @@ define inreg <48 x i16> @bitcast_v48f16_to_v48i16_scalar(<48 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB59_4
 ; GFX11-TRUE16-NEXT:  .LBB59_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v5.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v4.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v2.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v7, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v8, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v6, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v7, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v9, 16, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s15, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s14, s11
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s13, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s11, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s16, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s10, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s7, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s29 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s28 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s27 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s26 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s25 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s24 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v28, 0x200, s8 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v27, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v26, 0x200, s1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v25, 0x200, s2 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v24, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v23, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v29, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v26
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v25
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v24
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v28
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v10
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v50.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v51.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v52.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v53.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v54.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v55.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v5
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_5
 ; GFX11-TRUE16-NEXT:  .LBB59_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_2
 ; GFX11-TRUE16-NEXT:  .LBB59_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s29 :: v_dual_mov_b32 v15, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s27 :: v_dual_mov_b32 v17, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s25 :: v_dual_mov_b32 v10, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v12, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s21 :: v_dual_mov_b32 v6, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s19 :: v_dual_mov_b32 v8, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s17 :: v_dual_mov_b32 v29, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s3 :: v_dual_mov_b32 v25, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s1 :: v_dual_mov_b32 v27, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s45 :: v_dual_mov_b32 v31, s44
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s43 :: v_dual_mov_b32 v33, s42
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s41 :: v_dual_mov_b32 v35, s40
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, s15 :: v_dual_mov_b32 v37, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s12 :: v_dual_mov_b32 v39, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, s10 :: v_dual_mov_b32 v49, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, s8 :: v_dual_mov_b32 v51, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s5 :: v_dual_mov_b32 v53, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, s4 :: v_dual_mov_b32 v55, s14
 ; GFX11-TRUE16-NEXT:  .LBB59_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v31, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v33, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v38, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v49, 16, v51
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v30, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v24
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v20, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v39, 16, v26
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v27
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v52, 16, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v25
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v26
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v30, 16, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v28
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v3
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v48, 16, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v35, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v36, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v37, 16, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v32, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v53, 16, v64
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v49, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v48, 16, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v53
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v50, 16, v52
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v38, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v37, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v34, 16, v48
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v36, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v33, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v32, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v31, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v30, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v4
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v24 :: v_dual_mov_b32 v1, v25
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v26
 ; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v28 :: v_dual_mov_b32 v5, v29
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.832bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.832bit.ll
index fbee320c82c7f..97d040b545c09 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.832bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.832bit.ll
@@ -4133,341 +4133,173 @@ define inreg <52 x i16> @bitcast_v26i32_to_v52i16_scalar(<26 x i32> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr42
 ; GFX9-NEXT:    s_branch .LBB13_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v26i32_to_v52i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-TRUE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB13_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s11 :: v_dual_mov_b32 v19, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s9 :: v_dual_mov_b32 v21, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s7 :: v_dual_mov_b32 v23, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s6 :: v_dual_mov_b32 v25, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB13_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB13_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v26i32_to_v52i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-FAKE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-FAKE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB13_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB13_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    s_branch .LBB13_2
+; GFX11-LABEL: bitcast_v26i32_to_v52i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
+; GFX11-NEXT:    s_mov_b32 s78, 0
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB13_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s12, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
+; GFX11-NEXT:    s_cbranch_vccnz .LBB13_3
+; GFX11-NEXT:  .LBB13_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s10, s10, 3
+; GFX11-NEXT:    s_add_i32 s11, s11, 3
+; GFX11-NEXT:    s_add_i32 s9, s9, 3
+; GFX11-NEXT:    s_add_i32 s8, s8, 3
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s12, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s0, 16
+; GFX11-NEXT:  .LBB13_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s12
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB13_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    s_branch .LBB13_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -6427,184 +6259,118 @@ define inreg <26 x i32> @bitcast_v52i16_to_v26i32_scalar(<52 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v52i16_to_v26i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v7 :: v_dual_mov_b32 v33, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v3 :: v_dual_mov_b32 v37, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v1 :: v_dual_mov_b32 v39, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v7
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s28, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB15_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s17
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s15 :: v_dual_mov_b32 v17, s16
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB15_3
 ; GFX11-TRUE16-NEXT:  .LBB15_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v39
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v16, 16, v17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v22, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
@@ -8744,341 +8510,173 @@ define inreg <52 x half> @bitcast_v26i32_to_v52f16_scalar(<26 x i32> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr42
 ; GFX9-NEXT:    s_branch .LBB17_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v26i32_to_v52f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-TRUE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB17_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s11 :: v_dual_mov_b32 v19, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s9 :: v_dual_mov_b32 v21, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s7 :: v_dual_mov_b32 v23, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s6 :: v_dual_mov_b32 v25, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB17_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB17_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v26i32_to_v52f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-FAKE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-FAKE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB17_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB17_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    s_branch .LBB17_2
+; GFX11-LABEL: bitcast_v26i32_to_v52f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
+; GFX11-NEXT:    s_mov_b32 s78, 0
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB17_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s12, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
+; GFX11-NEXT:    s_cbranch_vccnz .LBB17_3
+; GFX11-NEXT:  .LBB17_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s10, s10, 3
+; GFX11-NEXT:    s_add_i32 s11, s11, 3
+; GFX11-NEXT:    s_add_i32 s9, s9, 3
+; GFX11-NEXT:    s_add_i32 s8, s8, 3
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s12, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s0, 16
+; GFX11-NEXT:  .LBB17_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s12
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB17_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    s_branch .LBB17_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -11321,184 +10919,118 @@ define inreg <26 x i32> @bitcast_v52f16_to_v26i32_scalar(<52 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v52f16_to_v26i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v7 :: v_dual_mov_b32 v33, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v3 :: v_dual_mov_b32 v37, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v1 :: v_dual_mov_b32 v39, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v7
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s28, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB19_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s17
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s15 :: v_dual_mov_b32 v17, s16
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB19_3
 ; GFX11-TRUE16-NEXT:  .LBB19_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v39
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v16, 16, v17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
@@ -14821,351 +14353,178 @@ define inreg <52 x i16> @bitcast_v26f32_to_v52i16_scalar(<26 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr38
 ; GFX9-NEXT:    s_branch .LBB29_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v26f32_to_v52i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v25, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v23, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v21, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v9, s22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s21 :: v_dual_mov_b32 v15, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v11, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-TRUE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:  .LBB29_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v64, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v55, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v53, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v52, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v51, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v50, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_and_b32 v4, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v2, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v66, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v65, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB29_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB29_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v26f32_to_v52i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v25, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v23, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v21, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v9, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s21 :: v_dual_mov_b32 v15, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v11, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-FAKE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:  .LBB29_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v64, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v55, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v53, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v52, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v51, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v50, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_and_b32 v4, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v2, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v66, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v65, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v32
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB29_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    s_branch .LBB29_2
+; GFX11-LABEL: bitcast_v26f32_to_v52i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v25, s1
+; GFX11-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v23, s3
+; GFX11-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v21, s17
+; GFX11-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v9, s22
+; GFX11-NEXT:    v_dual_mov_b32 v10, s21 :: v_dual_mov_b32 v15, s23
+; GFX11-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v11, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB29_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB29_3
+; GFX11-NEXT:  .LBB29_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
+; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
+; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v10, 1.0, v10
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
+; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
+; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:  .LBB29_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v8, v8, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v11, v64, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v12, v55, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v14, v53, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v15, v52, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v16, v51, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v50, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v6
+; GFX11-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_and_b32 v4, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
+; GFX11-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v2, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v33
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v9, v66, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v65, 16, v21
+; GFX11-NEXT:    v_lshl_or_b32 v13, v54, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v32
+; GFX11-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
+; GFX11-NEXT:    v_mov_b32_e32 v4, v28
+; GFX11-NEXT:    v_mov_b32_e32 v6, v30
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB29_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr8
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    s_branch .LBB29_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -17125,184 +16484,118 @@ define inreg <26 x float> @bitcast_v52i16_to_v26f32_scalar(<52 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v52i16_to_v26f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v7 :: v_dual_mov_b32 v33, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v3 :: v_dual_mov_b32 v37, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v1 :: v_dual_mov_b32 v39, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v7
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s28, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB31_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s17
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s15 :: v_dual_mov_b32 v17, s16
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB31_3
 ; GFX11-TRUE16-NEXT:  .LBB31_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v39
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v16, 16, v17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v22, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
@@ -19433,351 +18726,178 @@ define inreg <52 x half> @bitcast_v26f32_to_v52f16_scalar(<26 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr38
 ; GFX9-NEXT:    s_branch .LBB33_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v26f32_to_v52f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v25, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v23, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v21, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v9, s22
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s21 :: v_dual_mov_b32 v15, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v11, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-TRUE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:  .LBB33_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v64, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v55, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v53, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v52, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v51, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v50, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_and_b32 v4, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v2, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v66, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v65, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB33_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB33_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v26f32_to_v52f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v25, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v23, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v21, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v19, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v9, s22
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s21 :: v_dual_mov_b32 v15, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v11, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-FAKE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:  .LBB33_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v64, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v55, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v53, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v52, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v51, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v50, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_and_b32 v4, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v2, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v66, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v65, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v32
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB33_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr27
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    s_branch .LBB33_2
+; GFX11-LABEL: bitcast_v26f32_to_v52f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v25, s1
+; GFX11-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v23, s3
+; GFX11-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v21, s17
+; GFX11-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v19, s19
+; GFX11-NEXT:    v_dual_mov_b32 v18, s20 :: v_dual_mov_b32 v9, s22
+; GFX11-NEXT:    v_dual_mov_b32 v10, s21 :: v_dual_mov_b32 v15, s23
+; GFX11-NEXT:    v_dual_mov_b32 v14, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v12, s26 :: v_dual_mov_b32 v11, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB33_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB33_3
+; GFX11-NEXT:  .LBB33_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v12, 1.0, v12
+; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
+; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v10, 1.0, v10
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
+; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
+; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:  .LBB33_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v8, v8, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-NEXT:    v_lshl_or_b32 v27, v27, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v11, v64, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v12, v55, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v14, v53, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v15, v52, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v16, v51, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v50, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v6
+; GFX11-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_and_b32 v4, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
+; GFX11-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v2, 0xffff, v5
+; GFX11-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v33
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v9, v66, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v65, 16, v21
+; GFX11-NEXT:    v_lshl_or_b32 v13, v54, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v32
+; GFX11-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
+; GFX11-NEXT:    v_mov_b32_e32 v4, v28
+; GFX11-NEXT:    v_mov_b32_e32 v6, v30
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB33_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr27
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr8
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    s_branch .LBB33_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -22020,184 +21140,118 @@ define inreg <26 x float> @bitcast_v52f16_to_v26f32_scalar(<52 x half> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v52f16_to_v26f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v7 :: v_dual_mov_b32 v33, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v3 :: v_dual_mov_b32 v37, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v1 :: v_dual_mov_b32 v39, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v7
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s28, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB35_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s17
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s15 :: v_dual_mov_b32 v17, s16
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB35_3
 ; GFX11-TRUE16-NEXT:  .LBB35_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v39
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v16, 16, v17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
@@ -24730,341 +23784,173 @@ define inreg <52 x i16> @bitcast_v13i64_to_v52i16_scalar(<13 x i64> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr42
 ; GFX9-NEXT:    s_branch .LBB41_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v13i64_to_v52i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-TRUE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s8, s8, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s10, s10, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB41_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s11 :: v_dual_mov_b32 v19, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s9 :: v_dual_mov_b32 v21, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s7 :: v_dual_mov_b32 v23, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s6 :: v_dual_mov_b32 v25, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB41_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB41_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v13i64_to_v52i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-FAKE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-FAKE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s11, s11, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s10, s10, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s9, s9, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s7, s7, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB41_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB41_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    s_branch .LBB41_2
+; GFX11-LABEL: bitcast_v13i64_to_v52i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
+; GFX11-NEXT:    s_mov_b32 s78, 0
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB41_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s12, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
+; GFX11-NEXT:    s_cbranch_vccnz .LBB41_3
+; GFX11-NEXT:  .LBB41_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s11, s11, 3
+; GFX11-NEXT:    s_addc_u32 s10, s10, 0
+; GFX11-NEXT:    s_add_u32 s8, s8, 3
+; GFX11-NEXT:    s_addc_u32 s9, s9, 0
+; GFX11-NEXT:    s_add_u32 s6, s6, 3
+; GFX11-NEXT:    s_addc_u32 s7, s7, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s12, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s0, 16
+; GFX11-NEXT:  .LBB41_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s12
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB41_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    s_branch .LBB41_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -27024,184 +25910,118 @@ define inreg <13 x i64> @bitcast_v52i16_to_v13i64_scalar(<52 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v52i16_to_v13i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v7 :: v_dual_mov_b32 v33, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v3 :: v_dual_mov_b32 v37, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v1 :: v_dual_mov_b32 v39, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v7
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s28, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB43_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s17
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s15 :: v_dual_mov_b32 v17, s16
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB43_3
 ; GFX11-TRUE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v39
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v16, 16, v17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v22, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
@@ -29356,341 +28176,173 @@ define inreg <52 x half> @bitcast_v13i64_to_v52f16_scalar(<13 x i64> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr42
 ; GFX9-NEXT:    s_branch .LBB45_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v13i64_to_v52f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-TRUE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s8, s8, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s10, s10, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB45_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s11 :: v_dual_mov_b32 v19, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s9 :: v_dual_mov_b32 v21, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s7 :: v_dual_mov_b32 v23, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s6 :: v_dual_mov_b32 v25, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB45_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB45_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v13i64_to_v52f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v7
-; GFX11-FAKE16-NEXT:    s_mov_b32 s78, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-FAKE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s11, s11, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s10, s10, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s9, s9, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s7, s7, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB45_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s14
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s13
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB45_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    s_branch .LBB45_2
+; GFX11-LABEL: bitcast_v13i64_to_v52f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
+; GFX11-NEXT:    s_mov_b32 s78, 0
+; GFX11-NEXT:    s_and_b32 s12, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB45_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s12, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s78
+; GFX11-NEXT:    s_cbranch_vccnz .LBB45_3
+; GFX11-NEXT:  .LBB45_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s11, s11, 3
+; GFX11-NEXT:    s_addc_u32 s10, s10, 0
+; GFX11-NEXT:    s_add_u32 s8, s8, 3
+; GFX11-NEXT:    s_addc_u32 s9, s9, 0
+; GFX11-NEXT:    s_add_u32 s6, s6, 3
+; GFX11-NEXT:    s_addc_u32 s7, s7, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s12, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s0, 16
+; GFX11-NEXT:  .LBB45_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s14
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s13
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s12
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB45_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    s_branch .LBB45_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -31933,184 +30585,118 @@ define inreg <13 x i64> @bitcast_v52f16_to_v13i64_scalar(<52 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v52f16_to_v13i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v7 :: v_dual_mov_b32 v33, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v3 :: v_dual_mov_b32 v37, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v1 :: v_dual_mov_b32 v39, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v7
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s28, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB47_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s17
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s15 :: v_dual_mov_b32 v17, s16
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB47_3
 ; GFX11-TRUE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v39
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v16, 16, v17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
@@ -33713,351 +32299,178 @@ define inreg <52 x i16> @bitcast_v13f64_to_v52i16_scalar(<13 x double> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr25
 ; GFX9-NEXT:    s_branch .LBB49_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v13f64_to_v52i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v27, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v25, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v23, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v21, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v19, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s24 :: v_dual_mov_b32 v14, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s26 :: v_dual_mov_b32 v12, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:  .LBB49_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v65, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v64, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v66, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v53, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v52, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v55, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v51, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v50, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v30 :: v_dual_mov_b32 v7, v31
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v13f64_to_v52i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v27, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v25, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v23, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v21, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v19, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s24 :: v_dual_mov_b32 v14, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s26 :: v_dual_mov_b32 v12, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:  .LBB49_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v65, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v64, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v66, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v53, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v52, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v55, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v51, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v50, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v32
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, v30 :: v_dual_mov_b32 v7, v31
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
+; GFX11-LABEL: bitcast_v13f64_to_v52i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v27, s1
+; GFX11-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v25, s3
+; GFX11-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v23, s17
+; GFX11-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v21, s19
+; GFX11-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
+; GFX11-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v19, s23
+; GFX11-NEXT:    v_dual_mov_b32 v13, s24 :: v_dual_mov_b32 v14, s25
+; GFX11-NEXT:    v_dual_mov_b32 v11, s26 :: v_dual_mov_b32 v12, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_3
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
+; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
+; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
+; GFX11-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
+; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
+; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
+; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:  .LBB49_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v9, v65, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v64, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v27
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-NEXT:    v_lshl_or_b32 v27, v66, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v54, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v13, v53, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v14, v52, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v4
+; GFX11-NEXT:    v_lshl_or_b32 v11, v55, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v15, v15, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v6
+; GFX11-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v33
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v16, v51, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v50, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v32
+; GFX11-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
+; GFX11-NEXT:    v_mov_b32_e32 v4, v28
+; GFX11-NEXT:    v_dual_mov_b32 v6, v30 :: v_dual_mov_b32 v7, v31
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr8
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr15
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    s_branch .LBB49_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -36017,184 +34430,118 @@ define inreg <13 x double> @bitcast_v52i16_to_v13f64_scalar(<52 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v52i16_to_v13f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v7 :: v_dual_mov_b32 v33, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v3 :: v_dual_mov_b32 v37, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v1 :: v_dual_mov_b32 v39, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v7
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s28, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB51_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s17
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s15 :: v_dual_mov_b32 v17, s16
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB51_3
 ; GFX11-TRUE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v39
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v16, 16, v17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v22, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
@@ -38234,351 +36581,178 @@ define inreg <52 x half> @bitcast_v13f64_to_v52f16_scalar(<13 x double> inreg %a
 ; GFX9-NEXT:    ; implicit-def: $vgpr25
 ; GFX9-NEXT:    s_branch .LBB53_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v13f64_to_v52f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v27, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v25, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v23, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v21, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v19, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s24 :: v_dual_mov_b32 v14, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s26 :: v_dual_mov_b32 v12, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-TRUE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:  .LBB53_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v65, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v64, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v66, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v54, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v53, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v52, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v55, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v51, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v50, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v32
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v30 :: v_dual_mov_b32 v7, v31
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB53_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB53_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v13f64_to_v52f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v27, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v25, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v23, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v21, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v19, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s24 :: v_dual_mov_b32 v14, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s26 :: v_dual_mov_b32 v12, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-FAKE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:  .LBB53_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v65, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v64, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v66, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v54, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v53, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v52, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v55, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v51, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v50, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v32
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v28
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, v30 :: v_dual_mov_b32 v7, v31
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB53_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr28
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    s_branch .LBB53_2
+; GFX11-LABEL: bitcast_v13f64_to_v52f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-NEXT:    v_dual_mov_b32 v26, s0 :: v_dual_mov_b32 v27, s1
+; GFX11-NEXT:    v_dual_mov_b32 v24, s2 :: v_dual_mov_b32 v25, s3
+; GFX11-NEXT:    v_dual_mov_b32 v22, s16 :: v_dual_mov_b32 v23, s17
+; GFX11-NEXT:    v_dual_mov_b32 v20, s18 :: v_dual_mov_b32 v21, s19
+; GFX11-NEXT:    v_dual_mov_b32 v9, s20 :: v_dual_mov_b32 v10, s21
+; GFX11-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v19, s23
+; GFX11-NEXT:    v_dual_mov_b32 v13, s24 :: v_dual_mov_b32 v14, s25
+; GFX11-NEXT:    v_dual_mov_b32 v11, s26 :: v_dual_mov_b32 v12, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB53_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB53_3
+; GFX11-NEXT:  .LBB53_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
+; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
+; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
+; GFX11-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
+; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
+; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
+; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v28, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:  .LBB53_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
+; GFX11-NEXT:    v_lshl_or_b32 v9, v65, 16, v10
+; GFX11-NEXT:    v_lshl_or_b32 v10, v64, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v11
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v27
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-NEXT:    v_lshl_or_b32 v27, v66, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v30, v30, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v54, 16, v13
+; GFX11-NEXT:    v_lshl_or_b32 v13, v53, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v14, v52, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v19, v48, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v4
+; GFX11-NEXT:    v_lshl_or_b32 v11, v55, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v15, v15, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_lshl_or_b32 v18, v49, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v20, v39, 16, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v6
+; GFX11-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v4, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v28, v28, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v22, v37, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v33
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v26, v67, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v16, v51, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v50, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v21, v38, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v23, v36, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v24, v35, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v25, v34, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v32
+; GFX11-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
+; GFX11-NEXT:    v_mov_b32_e32 v4, v28
+; GFX11-NEXT:    v_dual_mov_b32 v6, v30 :: v_dual_mov_b32 v7, v31
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB53_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr28
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr8
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr15
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    s_branch .LBB53_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -40821,184 +38995,118 @@ define inreg <13 x double> @bitcast_v52f16_to_v13f64_scalar(<52 x half> inreg %a
 ; GFX11-TRUE16-LABEL: bitcast_v52f16_to_v13f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v7 :: v_dual_mov_b32 v33, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v5 :: v_dual_mov_b32 v35, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v3 :: v_dual_mov_b32 v37, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v1 :: v_dual_mov_b32 v39, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v7
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s40, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s28, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s29, s41
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB55_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s17
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s15 :: v_dual_mov_b32 v17, s16
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s40
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB55_3
 ; GFX11-TRUE16-NEXT:  .LBB55_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v39, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v38, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v37, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v36, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v35, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v34, 16, v50
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v33, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v32, 16, v48
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v39
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v16, 16, v17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v22 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
@@ -43612,25 +41720,41 @@ define inreg <52 x half> @bitcast_v52i16_to_v52f16_scalar(<52 x i16> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v52i16_to_v52f16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v25.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB57_3
@@ -43638,44 +41762,40 @@ define inreg <52 x half> @bitcast_v52i16_to_v52f16_scalar(<52 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB57_4
 ; GFX11-TRUE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v7.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v6.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v5.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v4.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v9, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v10, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v2.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v8, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v1.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v8, 16, v3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s15, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v9, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v10, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v11, 16, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s15, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s14, s11
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s12, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s11, s7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v25, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s16, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
@@ -43684,177 +41804,129 @@ define inreg <52 x half> @bitcast_v52i16_to_v52f16_scalar(<52 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, s8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s8, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s7, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s9
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s29, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s28, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s27, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s26, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s25, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s24, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, s6, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v32, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v29, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v28, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v27, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v26, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v52.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v53.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v54.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v55.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v64.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v65.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v66.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v67.l
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v30, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v31, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v26, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v33, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v32, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v29, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v28, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v27, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v28
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v26
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v31
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v30
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v7
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_5
 ; GFX11-TRUE16-NEXT:  .LBB57_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_2
 ; GFX11-TRUE16-NEXT:  .LBB57_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s29 :: v_dual_mov_b32 v17, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s27 :: v_dual_mov_b32 v12, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s25 :: v_dual_mov_b32 v14, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s23 :: v_dual_mov_b32 v8, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s21 :: v_dual_mov_b32 v10, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s19 :: v_dual_mov_b32 v31, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s17 :: v_dual_mov_b32 v27, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s3 :: v_dual_mov_b32 v29, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s1 :: v_dual_mov_b32 v33, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s45 :: v_dual_mov_b32 v35, s44
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, s43 :: v_dual_mov_b32 v37, s42
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s40 :: v_dual_mov_b32 v39, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, s14 :: v_dual_mov_b32 v49, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, s12 :: v_dual_mov_b32 v51, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s10 :: v_dual_mov_b32 v53, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, s8 :: v_dual_mov_b32 v55, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, s5 :: v_dual_mov_b32 v65, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v66, s4 :: v_dual_mov_b32 v67, s41
 ; GFX11-TRUE16-NEXT:  .LBB57_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v32
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v25, 16, v13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v7.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v19
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v27
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v67, 16, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v26
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v65, 16, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v54, 16, v67
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v64, 16, v28
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v55, 16, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v53, 16, v31
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v50, 16, v9
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v36, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v23, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v24, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v38, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v34, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v35, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v11
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v2.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v48, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v22, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v23, 16, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v4.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v5.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v50, 16, v32
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v39, 16, v28
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v49, 16, v52
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v6
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v32
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v26
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v37, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v2, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v52, 16, v54
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v66, 16, v68
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v48, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_and_b32 v4, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v51, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v49, 16, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v39, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v38, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v35, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v34, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v37, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v33
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v28
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: bitcast_v52i16_to_v52f16_scalar:
@@ -46046,25 +44118,41 @@ define inreg <52 x i16> @bitcast_v52f16_to_v52i16_scalar(<52 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v52f16_to_v52i16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v25.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v25.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB59_3
@@ -46072,44 +44160,40 @@ define inreg <52 x i16> @bitcast_v52f16_to_v52i16_scalar(<52 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB59_4
 ; GFX11-TRUE16-NEXT:  .LBB59_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v7.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v6.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v5.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v4.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v9, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v10, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v2.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v8, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v1.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s23, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s22, s14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v8, 16, v3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v9, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v10, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v11, 16, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s15, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s14, s11
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s12 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s12, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s11, s7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v25, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s16, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
@@ -46118,177 +44202,129 @@ define inreg <52 x i16> @bitcast_v52f16_to_v52i16_scalar(<52 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s8, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s7, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s9
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s29 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s28 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s27 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s26 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s25 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s24 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v32, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v29, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v28, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v27, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v26, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v52.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v53.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v54.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v55.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v64.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v65.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v66.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v67.l
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v30, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v31, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v26, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v33, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v32, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v29, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v28, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v27, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v28
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v26
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v31
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v30
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v7
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_5
 ; GFX11-TRUE16-NEXT:  .LBB59_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_2
 ; GFX11-TRUE16-NEXT:  .LBB59_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s10
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s29 :: v_dual_mov_b32 v17, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s27 :: v_dual_mov_b32 v12, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s25 :: v_dual_mov_b32 v14, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s23 :: v_dual_mov_b32 v8, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s21 :: v_dual_mov_b32 v10, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s19 :: v_dual_mov_b32 v31, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s17 :: v_dual_mov_b32 v27, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s3 :: v_dual_mov_b32 v29, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s1 :: v_dual_mov_b32 v33, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s45 :: v_dual_mov_b32 v35, s44
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, s43 :: v_dual_mov_b32 v37, s42
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s40 :: v_dual_mov_b32 v39, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, s14 :: v_dual_mov_b32 v49, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, s12 :: v_dual_mov_b32 v51, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s10 :: v_dual_mov_b32 v53, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, s8 :: v_dual_mov_b32 v55, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, s5 :: v_dual_mov_b32 v65, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v66, s4 :: v_dual_mov_b32 v67, s41
 ; GFX11-TRUE16-NEXT:  .LBB59_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v32
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v25, 16, v13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v7.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v19
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v27
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v67, 16, v33
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v26
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v65, 16, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v54, 16, v67
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v64, 16, v28
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v55, 16, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v53, 16, v31
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v50, 16, v9
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v51, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v36, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v23, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v24, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v38, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v34, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v35, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v11
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v2.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v48, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v22, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v23, 16, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v4.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v5.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v50, 16, v32
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v39, 16, v28
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v49, 16, v52
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v6
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v32
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v26
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v37, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v29 :: v_dual_and_b32 v2, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v52, 16, v54
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v66, 16, v68
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v48, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v36, 16, v50
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_and_b32 v4, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v51, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v49, 16, v55
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v39, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v38, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v35, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v34, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v37, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v4
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v33
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v26 :: v_dual_mov_b32 v3, v27
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v28
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: bitcast_v52f16_to_v52i16_scalar:
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.896bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.896bit.ll
index 462e50ac8412c..a43ce77b20631 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.896bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.896bit.ll
@@ -4433,367 +4433,186 @@ define inreg <56 x i16> @bitcast_v28i32_to_v56i16_scalar(<28 x i32> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr44
 ; GFX9-NEXT:    s_branch .LBB13_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v28i32_to_v56i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s90, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-TRUE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s13, s13, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB13_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s6 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB13_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB13_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v28i32_to_v56i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v9
-; GFX11-FAKE16-NEXT:    s_mov_b32 s90, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-FAKE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s13, s13, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB13_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB13_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    s_branch .LBB13_2
+; GFX11-LABEL: bitcast_v28i32_to_v56i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v9
+; GFX11-NEXT:    s_mov_b32 s90, 0
+; GFX11-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB13_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s14, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
+; GFX11-NEXT:    s_cbranch_vccnz .LBB13_3
+; GFX11-NEXT:  .LBB13_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s12, s12, 3
+; GFX11-NEXT:    s_add_i32 s13, s13, 3
+; GFX11-NEXT:    s_add_i32 s11, s11, 3
+; GFX11-NEXT:    s_add_i32 s10, s10, 3
+; GFX11-NEXT:    s_add_i32 s9, s9, 3
+; GFX11-NEXT:    s_add_i32 s8, s8, 3
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s14, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s0, 16
+; GFX11-NEXT:  .LBB13_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s13, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s12, s14
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
+; GFX11-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB13_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    s_branch .LBB13_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -6915,194 +6734,127 @@ define inreg <28 x i32> @bitcast_v56i16_to_v28i32_scalar(<56 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v28i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v9 :: v_dual_mov_b32 v33, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v7 :: v_dual_mov_b32 v35, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v5 :: v_dual_mov_b32 v37, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v3 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v1 :: v_dual_mov_b32 v49, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v9
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB15_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB15_3
 ; GFX11-TRUE16-NEXT:  .LBB15_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
@@ -9466,367 +9218,186 @@ define inreg <56 x half> @bitcast_v28i32_to_v56f16_scalar(<28 x i32> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr44
 ; GFX9-NEXT:    s_branch .LBB17_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v28i32_to_v56f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s90, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-TRUE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s13, s13, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB17_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s6 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB17_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB17_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v28i32_to_v56f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v9
-; GFX11-FAKE16-NEXT:    s_mov_b32 s90, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-FAKE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s13, s13, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB17_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB17_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    s_branch .LBB17_2
+; GFX11-LABEL: bitcast_v28i32_to_v56f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v9
+; GFX11-NEXT:    s_mov_b32 s90, 0
+; GFX11-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB17_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s14, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
+; GFX11-NEXT:    s_cbranch_vccnz .LBB17_3
+; GFX11-NEXT:  .LBB17_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s12, s12, 3
+; GFX11-NEXT:    s_add_i32 s13, s13, 3
+; GFX11-NEXT:    s_add_i32 s11, s11, 3
+; GFX11-NEXT:    s_add_i32 s10, s10, 3
+; GFX11-NEXT:    s_add_i32 s9, s9, 3
+; GFX11-NEXT:    s_add_i32 s8, s8, 3
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s14, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s0, 16
+; GFX11-NEXT:  .LBB17_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s13, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s12, s14
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
+; GFX11-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB17_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    s_branch .LBB17_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -12269,194 +11840,127 @@ define inreg <28 x i32> @bitcast_v56f16_to_v28i32_scalar(<56 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v28i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v9 :: v_dual_mov_b32 v33, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v7 :: v_dual_mov_b32 v35, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v5 :: v_dual_mov_b32 v37, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v3 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v1 :: v_dual_mov_b32 v49, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v9
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB19_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB19_3
 ; GFX11-TRUE16-NEXT:  .LBB19_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
@@ -16039,377 +15543,191 @@ define inreg <56 x i16> @bitcast_v28f32_to_v56i16_scalar(<28 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    s_branch .LBB29_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v28f32_to_v56i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s0 :: v_dual_mov_b32 v27, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s2 :: v_dual_mov_b32 v25, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s16 :: v_dual_mov_b32 v23, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s18 :: v_dual_mov_b32 v21, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v19, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v11, s24
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s23 :: v_dual_mov_b32 v17, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s26 :: v_dual_mov_b32 v15, s28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s27 :: v_dual_mov_b32 v13, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-TRUE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-TRUE16-NEXT:  .LBB29_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v67, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v65, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v66, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v35 :: v_dual_and_b32 v0, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v71, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v69, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v68, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v64, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v3, v33
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v34
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v36
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v28 :: v_dual_mov_b32 v9, v29
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB29_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB29_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v28f32_to_v56i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s0 :: v_dual_mov_b32 v27, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s2 :: v_dual_mov_b32 v25, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s16 :: v_dual_mov_b32 v23, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s18 :: v_dual_mov_b32 v21, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v19, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v11, s24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s23 :: v_dual_mov_b32 v15, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v13, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-FAKE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-FAKE16-NEXT:  .LBB29_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v65, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v64, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v69, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v68, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v66, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v5, v35
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v71, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_and_b32 v2, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v67, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v30
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v3, v33
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v34
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v6, v36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, v28 :: v_dual_mov_b32 v9, v29
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB29_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    s_branch .LBB29_2
+; GFX11-LABEL: bitcast_v28f32_to_v56i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX11-NEXT:    v_dual_mov_b32 v28, s0 :: v_dual_mov_b32 v27, s1
+; GFX11-NEXT:    v_dual_mov_b32 v26, s2 :: v_dual_mov_b32 v25, s3
+; GFX11-NEXT:    v_dual_mov_b32 v24, s16 :: v_dual_mov_b32 v23, s17
+; GFX11-NEXT:    v_dual_mov_b32 v22, s18 :: v_dual_mov_b32 v21, s19
+; GFX11-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v19, s21
+; GFX11-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v11, s24
+; GFX11-NEXT:    v_dual_mov_b32 v12, s23 :: v_dual_mov_b32 v15, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v13, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB29_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB29_3
+; GFX11-NEXT:  .LBB29_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
+; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
+; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v12, 1.0, v12
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
+; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
+; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
+; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
+; GFX11-NEXT:  .LBB29_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v27
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-NEXT:    v_lshl_or_b32 v35, v35, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v37, v37, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v11
+; GFX11-NEXT:    v_lshl_or_b32 v10, v10, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v16, v65, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v64, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v69, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v13, v68, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v15, v66, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v8
+; GFX11-NEXT:    v_mov_b32_e32 v5, v35
+; GFX11-NEXT:    v_lshl_or_b32 v34, v34, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v31
+; GFX11-NEXT:    v_lshl_or_b32 v30, v30, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v28, v71, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
+; GFX11-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_and_b32 v2, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v36, v36, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v9
+; GFX11-NEXT:    v_lshl_or_b32 v14, v67, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v30
+; GFX11-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v3, v33
+; GFX11-NEXT:    v_mov_b32_e32 v4, v34
+; GFX11-NEXT:    v_mov_b32_e32 v6, v36
+; GFX11-NEXT:    v_dual_mov_b32 v8, v28 :: v_dual_mov_b32 v9, v29
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB29_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr71
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr10
+; GFX11-NEXT:    ; implicit-def: $vgpr70
+; GFX11-NEXT:    ; implicit-def: $vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr68
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    s_branch .LBB29_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -18531,194 +17849,127 @@ define inreg <28 x float> @bitcast_v56i16_to_v28f32_scalar(<56 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v28f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v9 :: v_dual_mov_b32 v33, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v7 :: v_dual_mov_b32 v35, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v5 :: v_dual_mov_b32 v37, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v3 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v1 :: v_dual_mov_b32 v49, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v9
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB31_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB31_3
 ; GFX11-TRUE16-NEXT:  .LBB31_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
@@ -21062,377 +20313,191 @@ define inreg <56 x half> @bitcast_v28f32_to_v56f16_scalar(<28 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr50
 ; GFX9-NEXT:    s_branch .LBB33_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v28f32_to_v56f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s0 :: v_dual_mov_b32 v27, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s2 :: v_dual_mov_b32 v25, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s16 :: v_dual_mov_b32 v23, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s18 :: v_dual_mov_b32 v21, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v19, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v11, s24
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s23 :: v_dual_mov_b32 v17, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s26 :: v_dual_mov_b32 v15, s28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s27 :: v_dual_mov_b32 v13, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-TRUE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-TRUE16-NEXT:  .LBB33_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v67, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v65, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v66, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v35 :: v_dual_and_b32 v0, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v71, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v69, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v68, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v64, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v3, v33
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v34
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v36
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v28 :: v_dual_mov_b32 v9, v29
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB33_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB33_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v28f32_to_v56f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s0 :: v_dual_mov_b32 v27, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s2 :: v_dual_mov_b32 v25, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s16 :: v_dual_mov_b32 v23, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s18 :: v_dual_mov_b32 v21, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v19, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v11, s24
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s23 :: v_dual_mov_b32 v15, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v13, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-FAKE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v12, 1.0, v12
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
-; GFX11-FAKE16-NEXT:  .LBB33_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v65, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v64, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v69, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v68, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v66, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v5, v35
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v71, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_and_b32 v2, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v67, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v30
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v3, v33
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v34
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v6, v36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, v28 :: v_dual_mov_b32 v9, v29
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB33_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    s_branch .LBB33_2
+; GFX11-LABEL: bitcast_v28f32_to_v56f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX11-NEXT:    v_dual_mov_b32 v28, s0 :: v_dual_mov_b32 v27, s1
+; GFX11-NEXT:    v_dual_mov_b32 v26, s2 :: v_dual_mov_b32 v25, s3
+; GFX11-NEXT:    v_dual_mov_b32 v24, s16 :: v_dual_mov_b32 v23, s17
+; GFX11-NEXT:    v_dual_mov_b32 v22, s18 :: v_dual_mov_b32 v21, s19
+; GFX11-NEXT:    v_dual_mov_b32 v20, s20 :: v_dual_mov_b32 v19, s21
+; GFX11-NEXT:    v_dual_mov_b32 v18, s22 :: v_dual_mov_b32 v11, s24
+; GFX11-NEXT:    v_dual_mov_b32 v12, s23 :: v_dual_mov_b32 v15, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v13, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB33_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB33_3
+; GFX11-NEXT:  .LBB33_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
+; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v14, 1.0, v14
+; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v12, 1.0, v12
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
+; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
+; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
+; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
+; GFX11-NEXT:  .LBB33_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v27
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-NEXT:    v_lshl_or_b32 v35, v35, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v37, v37, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v11
+; GFX11-NEXT:    v_lshl_or_b32 v10, v10, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v16, v65, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v64, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v69, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v13, v68, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v15, v66, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v8
+; GFX11-NEXT:    v_mov_b32_e32 v5, v35
+; GFX11-NEXT:    v_lshl_or_b32 v34, v34, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v31
+; GFX11-NEXT:    v_lshl_or_b32 v30, v30, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v28, v71, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
+; GFX11-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_and_b32 v2, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v36, v36, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v9
+; GFX11-NEXT:    v_lshl_or_b32 v14, v67, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v30
+; GFX11-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v3, v33
+; GFX11-NEXT:    v_mov_b32_e32 v4, v34
+; GFX11-NEXT:    v_mov_b32_e32 v6, v36
+; GFX11-NEXT:    v_dual_mov_b32 v8, v28 :: v_dual_mov_b32 v9, v29
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB33_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr71
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr10
+; GFX11-NEXT:    ; implicit-def: $vgpr70
+; GFX11-NEXT:    ; implicit-def: $vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr68
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    s_branch .LBB33_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -23875,194 +22940,127 @@ define inreg <28 x float> @bitcast_v56f16_to_v28f32_scalar(<56 x half> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v28f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v9 :: v_dual_mov_b32 v33, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v7 :: v_dual_mov_b32 v35, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v5 :: v_dual_mov_b32 v37, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v3 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v1 :: v_dual_mov_b32 v49, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v9
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB35_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB35_3
 ; GFX11-TRUE16-NEXT:  .LBB35_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
@@ -26806,367 +25804,186 @@ define inreg <56 x i16> @bitcast_v14i64_to_v56i16_scalar(<14 x i64> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr44
 ; GFX9-NEXT:    s_branch .LBB41_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v14i64_to_v56i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s90, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-TRUE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s8, s8, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s10, s10, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s13, s13, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s12, s12, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB41_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s6 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB41_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB41_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v14i64_to_v56i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v9
-; GFX11-FAKE16-NEXT:    s_mov_b32 s90, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-FAKE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s13, s13, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s12, s12, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s11, s11, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s9, s9, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s7, s7, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB41_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB41_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    s_branch .LBB41_2
+; GFX11-LABEL: bitcast_v14i64_to_v56i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v9
+; GFX11-NEXT:    s_mov_b32 s90, 0
+; GFX11-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB41_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s14, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
+; GFX11-NEXT:    s_cbranch_vccnz .LBB41_3
+; GFX11-NEXT:  .LBB41_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s13, s13, 3
+; GFX11-NEXT:    s_addc_u32 s12, s12, 0
+; GFX11-NEXT:    s_add_u32 s10, s10, 3
+; GFX11-NEXT:    s_addc_u32 s11, s11, 0
+; GFX11-NEXT:    s_add_u32 s8, s8, 3
+; GFX11-NEXT:    s_addc_u32 s9, s9, 0
+; GFX11-NEXT:    s_add_u32 s6, s6, 3
+; GFX11-NEXT:    s_addc_u32 s7, s7, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s14, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s0, 16
+; GFX11-NEXT:  .LBB41_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s13, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s12, s14
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
+; GFX11-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB41_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    s_branch .LBB41_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -29288,194 +28105,127 @@ define inreg <14 x i64> @bitcast_v56i16_to_v14i64_scalar(<56 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v14i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v9 :: v_dual_mov_b32 v33, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v7 :: v_dual_mov_b32 v35, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v5 :: v_dual_mov_b32 v37, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v3 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v1 :: v_dual_mov_b32 v49, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v9
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB43_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB43_3
 ; GFX11-TRUE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
@@ -31853,367 +30603,186 @@ define inreg <56 x half> @bitcast_v14i64_to_v56f16_scalar(<14 x i64> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr44
 ; GFX9-NEXT:    s_branch .LBB45_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v14i64_to_v56f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s90, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-TRUE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s8, s8, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s10, s10, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s13, s13, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s12, s12, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB45_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s13 :: v_dual_mov_b32 v19, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s11 :: v_dual_mov_b32 v21, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s9 :: v_dual_mov_b32 v23, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s7 :: v_dual_mov_b32 v25, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s6 :: v_dual_mov_b32 v27, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB45_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr15_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB45_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v14i64_to_v56f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v9
-; GFX11-FAKE16-NEXT:    s_mov_b32 s90, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-FAKE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s13, s13, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s12, s12, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s11, s11, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s9, s9, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s7, s7, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s15, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB45_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s40
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s15
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB45_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr15
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    s_branch .LBB45_2
+; GFX11-LABEL: bitcast_v14i64_to_v56f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v9
+; GFX11-NEXT:    s_mov_b32 s90, 0
+; GFX11-NEXT:    s_and_b32 s14, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB45_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s14, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s90
+; GFX11-NEXT:    s_cbranch_vccnz .LBB45_3
+; GFX11-NEXT:  .LBB45_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s13, s13, 3
+; GFX11-NEXT:    s_addc_u32 s12, s12, 0
+; GFX11-NEXT:    s_add_u32 s10, s10, 3
+; GFX11-NEXT:    s_addc_u32 s11, s11, 0
+; GFX11-NEXT:    s_add_u32 s8, s8, 3
+; GFX11-NEXT:    s_addc_u32 s9, s9, 0
+; GFX11-NEXT:    s_add_u32 s6, s6, 3
+; GFX11-NEXT:    s_addc_u32 s7, s7, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s14, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s15, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s40, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s0, 16
+; GFX11-NEXT:  .LBB45_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s89
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s88
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s79
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s78
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s40
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s13, s15
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s12, s14
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
+; GFX11-NEXT:    v_dual_mov_b32 v26, s13 :: v_dual_mov_b32 v27, s12
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB45_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    ; implicit-def: $sgpr15
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    s_branch .LBB45_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -34656,194 +33225,127 @@ define inreg <14 x i64> @bitcast_v56f16_to_v14i64_scalar(<56 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v14i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v9 :: v_dual_mov_b32 v33, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v7 :: v_dual_mov_b32 v35, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v5 :: v_dual_mov_b32 v37, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v3 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v1 :: v_dual_mov_b32 v49, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v9
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB47_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB47_3
 ; GFX11-TRUE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
@@ -36605,375 +35107,191 @@ define inreg <56 x i16> @bitcast_v14f64_to_v56i16_scalar(<14 x double> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr27
 ; GFX9-NEXT:    s_branch .LBB49_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v14f64_to_v56i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v27, s0 :: v_dual_mov_b32 v28, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v25, s2 :: v_dual_mov_b32 v26, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v23, s16 :: v_dual_mov_b32 v24, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v21, s18 :: v_dual_mov_b32 v22, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v19, s20 :: v_dual_mov_b32 v20, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s24 :: v_dual_mov_b32 v18, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v16, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s28 :: v_dual_mov_b32 v14, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
-; GFX11-TRUE16-NEXT:  .LBB49_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v71, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, v29 :: v_dual_and_b32 v4, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v69, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v68, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v64, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v36 :: v_dual_and_b32 v1, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v67, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v66, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v65, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v30 :: v_dual_mov_b32 v1, v31
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v5, v35
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_mov_b32 v8, v28
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v14f64_to_v56i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s0 :: v_dual_mov_b32 v28, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, s2 :: v_dual_mov_b32 v26, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s16 :: v_dual_mov_b32 v24, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v21, s18 :: v_dual_mov_b32 v22, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s20 :: v_dual_mov_b32 v20, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s24 :: v_dual_mov_b32 v18, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s26 :: v_dual_mov_b32 v14, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
-; GFX11-FAKE16-NEXT:  .LBB49_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v71, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v66, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v6, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v9, v29
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v69, 16, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v68, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v64, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v34
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v67, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v65, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, v30 :: v_dual_mov_b32 v1, v31
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v5, v35
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_mov_b32 v8, v28
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
+; GFX11-LABEL: bitcast_v14f64_to_v56i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX11-NEXT:    v_dual_mov_b32 v27, s0 :: v_dual_mov_b32 v28, s1
+; GFX11-NEXT:    v_dual_mov_b32 v25, s2 :: v_dual_mov_b32 v26, s3
+; GFX11-NEXT:    v_dual_mov_b32 v23, s16 :: v_dual_mov_b32 v24, s17
+; GFX11-NEXT:    v_dual_mov_b32 v21, s18 :: v_dual_mov_b32 v22, s19
+; GFX11-NEXT:    v_dual_mov_b32 v19, s20 :: v_dual_mov_b32 v20, s21
+; GFX11-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
+; GFX11-NEXT:    v_dual_mov_b32 v17, s24 :: v_dual_mov_b32 v18, s25
+; GFX11-NEXT:    v_dual_mov_b32 v13, s26 :: v_dual_mov_b32 v14, s27
+; GFX11-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_3
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
+; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
+; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
+; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
+; GFX11-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
+; GFX11-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
+; GFX11-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
+; GFX11-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
+; GFX11-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
+; GFX11-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
+; GFX11-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
+; GFX11-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
+; GFX11-NEXT:  .LBB49_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v37, v37, 16, v22
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v28, v71, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v36, v36, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v25
+; GFX11-NEXT:    v_lshl_or_b32 v10, v10, 16, v11
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v16
+; GFX11-NEXT:    v_lshl_or_b32 v15, v66, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v6
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v9
+; GFX11-NEXT:    v_mov_b32_e32 v6, v36
+; GFX11-NEXT:    v_lshl_or_b32 v34, v34, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v26
+; GFX11-NEXT:    v_mov_b32_e32 v9, v29
+; GFX11-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v69, 16, v17
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v18
+; GFX11-NEXT:    v_lshl_or_b32 v30, v30, 16, v27
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v13, v68, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v64, 16, v21
+; GFX11-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v4, v34
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v14, v67, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v16, v65, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, v33
+; GFX11-NEXT:    v_lshl_or_b32 v35, v35, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
+; GFX11-NEXT:    v_dual_mov_b32 v0, v30 :: v_dual_mov_b32 v1, v31
+; GFX11-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v5, v35
+; GFX11-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_mov_b32 v8, v28
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr71
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr10
+; GFX11-NEXT:    ; implicit-def: $vgpr70
+; GFX11-NEXT:    ; implicit-def: $vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr68
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    s_branch .LBB49_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -39095,194 +37413,127 @@ define inreg <14 x double> @bitcast_v56i16_to_v14f64_scalar(<56 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v14f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v9 :: v_dual_mov_b32 v33, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v7 :: v_dual_mov_b32 v35, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v5 :: v_dual_mov_b32 v37, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v3 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v1 :: v_dual_mov_b32 v49, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v9
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB51_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB51_3
 ; GFX11-TRUE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
@@ -41539,375 +39790,191 @@ define inreg <56 x half> @bitcast_v14f64_to_v56f16_scalar(<14 x double> inreg %a
 ; GFX9-NEXT:    ; implicit-def: $vgpr27
 ; GFX9-NEXT:    s_branch .LBB53_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v14f64_to_v56f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v27, s0 :: v_dual_mov_b32 v28, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v25, s2 :: v_dual_mov_b32 v26, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v23, s16 :: v_dual_mov_b32 v24, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v21, s18 :: v_dual_mov_b32 v22, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v19, s20 :: v_dual_mov_b32 v20, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s24 :: v_dual_mov_b32 v18, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s26 :: v_dual_mov_b32 v16, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s28 :: v_dual_mov_b32 v14, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-TRUE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
-; GFX11-TRUE16-NEXT:  .LBB53_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v71, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, v29 :: v_dual_and_b32 v4, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v69, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v68, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v64, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v36 :: v_dual_and_b32 v1, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v67, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v66, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v65, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v30 :: v_dual_mov_b32 v1, v31
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v5, v35
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_mov_b32 v8, v28
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB53_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB53_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v14f64_to_v56f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v27, s0 :: v_dual_mov_b32 v28, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, s2 :: v_dual_mov_b32 v26, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v23, s16 :: v_dual_mov_b32 v24, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v21, s18 :: v_dual_mov_b32 v22, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v19, s20 :: v_dual_mov_b32 v20, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s24 :: v_dual_mov_b32 v18, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s26 :: v_dual_mov_b32 v14, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-FAKE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
-; GFX11-FAKE16-NEXT:  .LBB53_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v71, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v11
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v66, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v6, v36
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v9, v29
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v69, 16, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v68, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v64, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v34
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v67, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v65, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, v30 :: v_dual_mov_b32 v1, v31
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v5, v35
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_mov_b32 v8, v28
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB53_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr30
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr29
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    s_branch .LBB53_2
+; GFX11-LABEL: bitcast_v14f64_to_v56f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX11-NEXT:    v_dual_mov_b32 v27, s0 :: v_dual_mov_b32 v28, s1
+; GFX11-NEXT:    v_dual_mov_b32 v25, s2 :: v_dual_mov_b32 v26, s3
+; GFX11-NEXT:    v_dual_mov_b32 v23, s16 :: v_dual_mov_b32 v24, s17
+; GFX11-NEXT:    v_dual_mov_b32 v21, s18 :: v_dual_mov_b32 v22, s19
+; GFX11-NEXT:    v_dual_mov_b32 v19, s20 :: v_dual_mov_b32 v20, s21
+; GFX11-NEXT:    v_dual_mov_b32 v11, s22 :: v_dual_mov_b32 v12, s23
+; GFX11-NEXT:    v_dual_mov_b32 v17, s24 :: v_dual_mov_b32 v18, s25
+; GFX11-NEXT:    v_dual_mov_b32 v13, s26 :: v_dual_mov_b32 v14, s27
+; GFX11-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB53_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB53_3
+; GFX11-NEXT:  .LBB53_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
+; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
+; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
+; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
+; GFX11-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
+; GFX11-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
+; GFX11-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
+; GFX11-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
+; GFX11-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
+; GFX11-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
+; GFX11-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
+; GFX11-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
+; GFX11-NEXT:  .LBB53_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v37, v37, 16, v22
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v28, v71, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v14
+; GFX11-NEXT:    v_lshl_or_b32 v29, v29, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-NEXT:    v_lshl_or_b32 v36, v36, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v25
+; GFX11-NEXT:    v_lshl_or_b32 v10, v10, 16, v11
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v16
+; GFX11-NEXT:    v_lshl_or_b32 v15, v66, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v19, v54, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v22, v51, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v6
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v9
+; GFX11-NEXT:    v_mov_b32_e32 v6, v36
+; GFX11-NEXT:    v_lshl_or_b32 v34, v34, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v26
+; GFX11-NEXT:    v_mov_b32_e32 v9, v29
+; GFX11-NEXT:    v_lshl_or_b32 v11, v70, 16, v12
+; GFX11-NEXT:    v_lshl_or_b32 v12, v69, 16, v17
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v18
+; GFX11-NEXT:    v_lshl_or_b32 v30, v30, 16, v27
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-NEXT:    v_lshl_or_b32 v13, v68, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v64, 16, v21
+; GFX11-NEXT:    v_lshl_or_b32 v21, v52, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v27, v38, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v4, v34
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v14, v67, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v16, v65, 16, v20
+; GFX11-NEXT:    v_lshl_or_b32 v18, v55, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v20, v53, 16, v2
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v7
+; GFX11-NEXT:    v_lshl_or_b32 v26, v39, 16, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, v33
+; GFX11-NEXT:    v_lshl_or_b32 v35, v35, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v23, v50, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v24, v49, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v25, v48, 16, v2
+; GFX11-NEXT:    v_dual_mov_b32 v0, v30 :: v_dual_mov_b32 v1, v31
+; GFX11-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v5, v35
+; GFX11-NEXT:    v_dual_mov_b32 v7, v37 :: v_dual_mov_b32 v8, v28
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB53_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr30
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr71
+; GFX11-NEXT:    ; implicit-def: $vgpr29
+; GFX11-NEXT:    ; implicit-def: $vgpr10
+; GFX11-NEXT:    ; implicit-def: $vgpr70
+; GFX11-NEXT:    ; implicit-def: $vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr68
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    s_branch .LBB53_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -44350,194 +42417,127 @@ define inreg <14 x double> @bitcast_v56f16_to_v14f64_scalar(<56 x half> inreg %a
 ; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v14f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v9 :: v_dual_mov_b32 v33, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v7 :: v_dual_mov_b32 v35, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v5 :: v_dual_mov_b32 v37, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v3 :: v_dual_mov_b32 v39, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v1 :: v_dual_mov_b32 v49, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v50, 0xffff, v9
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB55_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s59 :: v_dual_mov_b32 v5, s60
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s61 :: v_dual_mov_b32 v7, s62
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s63 :: v_dual_mov_b32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB55_3
 ; GFX11-TRUE16-NEXT:  .LBB55_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v49, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v48, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v39, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v38, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v36, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v35, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v34, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v33, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v32, 16, v50
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s42
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s40
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s18
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s14
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s22
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s11
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s10
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v31, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s28
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
@@ -47438,25 +45438,45 @@ define inreg <56 x half> @bitcast_v56i16_to_v56f16_scalar(<56 x i16> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v56i16_to_v56f16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v27.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB57_3
@@ -47464,50 +45484,44 @@ define inreg <56 x half> @bitcast_v56i16_to_v56f16_scalar(<56 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB57_4
 ; GFX11-TRUE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v9.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v8.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v7.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v6.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v5.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v11, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v12, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v10, 16, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v4.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v2.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v13, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v10, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v3.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v27, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v26, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v25, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v11, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v12, 16, v1
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v10, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v13, 16, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s14, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s13, s9
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, s9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s10, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s9, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s16, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
@@ -47518,188 +45532,138 @@ define inreg <56 x half> @bitcast_v56i16_to_v56f16_scalar(<56 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, s8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s8, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s7, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s11
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s29, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s28, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s26, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s25, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s24, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v22, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v28, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v29, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v33, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v34, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v35, s8, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v32, s0, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v31, s1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v30, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v33, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v28, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v54.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v55.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v64.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v65.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v66.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v67.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v68.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v69.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v70.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v71.l
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v37, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v36, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v31
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v30
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v28
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v9
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_5
 ; GFX11-TRUE16-NEXT:  .LBB57_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_2
 ; GFX11-TRUE16-NEXT:  .LBB57_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s29 :: v_dual_mov_b32 v14, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v16, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s25 :: v_dual_mov_b32 v10, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v12, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s21 :: v_dual_mov_b32 v29, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, s19 :: v_dual_mov_b32 v34, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, s17 :: v_dual_mov_b32 v36, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, s3 :: v_dual_mov_b32 v30, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, s1 :: v_dual_mov_b32 v32, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s45 :: v_dual_mov_b32 v39, s44
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, s42 :: v_dual_mov_b32 v49, s41
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, s40 :: v_dual_mov_b32 v51, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s14 :: v_dual_mov_b32 v53, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, s12 :: v_dual_mov_b32 v55, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, s10 :: v_dual_mov_b32 v65, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v66, s8 :: v_dual_mov_b32 v67, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v68, s5 :: v_dual_mov_b32 v69, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v70, s4 :: v_dual_mov_b32 v71, s43
 ; GFX11-TRUE16-NEXT:  .LBB57_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v34, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v35, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v38, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v36, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v23, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v24, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v26, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v27, 16, v19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v5.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v7.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v9.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v25, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v37, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v50, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v10
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v39, 16, v12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v49, 16, v20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v52, 16, v31
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v70, 16, v31
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v68, 16, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v80, 0xffff, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v6.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v8.h
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v0, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v66, 16, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v3
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v51, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v2, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v71, 16, v32
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v53, 16, v54
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v30 :: v_dual_mov_b32 v5, v35
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v36
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v7, v37
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v48, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v69, 16, v80
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v64, 16, v70
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v35 :: v_dual_and_b32 v0, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v67, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v65, 16, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v55, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v54, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v53, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v52, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v51, 16, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v50, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v49, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v48, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v39, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v38, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v4
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v30
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v32
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v34
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v36 :: v_dual_mov_b32 v7, v37
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v28 :: v_dual_mov_b32 v9, v29
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: bitcast_v56i16_to_v56f16_scalar:
@@ -50104,25 +48068,45 @@ define inreg <56 x i16> @bitcast_v56f16_to_v56i16_scalar(<56 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v56f16_to_v56i16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v10
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v27.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v27.h
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB59_3
@@ -50130,50 +48114,44 @@ define inreg <56 x i16> @bitcast_v56f16_to_v56i16_scalar(<56 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB59_4
 ; GFX11-TRUE16-NEXT:  .LBB59_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v9.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v8.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v7.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v6.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v5.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v11, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v12, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v10, 16, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v4.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v2.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v1.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v13, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v10, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v3.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v27, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v26, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v25, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v11, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v12, 16, v1
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v10, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v13, 16, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s14, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s13, s9
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s10, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s9, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s16, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
@@ -50184,188 +48162,138 @@ define inreg <56 x i16> @bitcast_v56f16_to_v56i16_scalar(<56 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s8, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s7, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s15
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s11
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s29 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s28 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s26 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s25 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s24 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v22, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v28, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v29, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v33, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v34, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v35, 0x200, s8 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v32, 0x200, s0 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v31, 0x200, s1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v30, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v33, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v28, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v54.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v55.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v64.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v65.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v66.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v67.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v68.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v69.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v70.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v71.l
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v37, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v36, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v31
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v30
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v28
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v9
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_5
 ; GFX11-TRUE16-NEXT:  .LBB59_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_2
 ; GFX11-TRUE16-NEXT:  .LBB59_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s12
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s29 :: v_dual_mov_b32 v14, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s27 :: v_dual_mov_b32 v16, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s25 :: v_dual_mov_b32 v10, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v12, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s21 :: v_dual_mov_b32 v29, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, s19 :: v_dual_mov_b32 v34, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, s17 :: v_dual_mov_b32 v36, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, s3 :: v_dual_mov_b32 v30, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, s1 :: v_dual_mov_b32 v32, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s45 :: v_dual_mov_b32 v39, s44
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, s42 :: v_dual_mov_b32 v49, s41
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, s40 :: v_dual_mov_b32 v51, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s14 :: v_dual_mov_b32 v53, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, s12 :: v_dual_mov_b32 v55, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, s10 :: v_dual_mov_b32 v65, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v66, s8 :: v_dual_mov_b32 v67, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v68, s5 :: v_dual_mov_b32 v69, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v70, s4 :: v_dual_mov_b32 v71, s43
 ; GFX11-TRUE16-NEXT:  .LBB59_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v34, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v35, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v29, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v38, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v36, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v23, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v24, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v26, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v27, 16, v19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v5.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v7.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v9.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v25, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v34
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v37, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v50, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v10
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v39, 16, v12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v49, 16, v20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v52, 16, v31
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v70, 16, v31
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v68, 16, v37
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v80, 0xffff, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
 ; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v6.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v8.h
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v0, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v66, 16, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v3
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v51, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v2, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v71, 16, v32
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v53, 16, v54
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v30 :: v_dual_mov_b32 v5, v35
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v36
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v32 :: v_dual_mov_b32 v7, v37
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v48, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v69, 16, v80
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v34
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v64, 16, v70
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v28
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, v35 :: v_dual_and_b32 v0, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v67, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v65, 16, v69
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v55, 16, v29
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v54, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v53, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v52, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v51, 16, v65
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v51, 0xffff, v14
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v50, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v49, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v48, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v39, 16, v51
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v38, 16, v52
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v4
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v30
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v32
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v34
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v36 :: v_dual_mov_b32 v7, v37
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v28 :: v_dual_mov_b32 v9, v29
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: bitcast_v56f16_to_v56i16_scalar:
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
index 94ed6276bd051..4f46875076809 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.960bit.ll
@@ -4722,393 +4722,199 @@ define inreg <60 x i16> @bitcast_v30i32_to_v60i16_scalar(<30 x i32> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr46
 ; GFX9-NEXT:    s_branch .LBB13_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v30i32_to_v60i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s94, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s15, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-TRUE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s13, s13, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s15, s15, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s15, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB13_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s15, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s14, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s40
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s15 :: v_dual_mov_b32 v19, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s13 :: v_dual_mov_b32 v21, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s11 :: v_dual_mov_b32 v23, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s9 :: v_dual_mov_b32 v25, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s7 :: v_dual_mov_b32 v27, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s6 :: v_dual_mov_b32 v29, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB13_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB13_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v30i32_to_v60i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v11
-; GFX11-FAKE16-NEXT:    s_mov_b32 s94, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB13_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s15, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB13_3
-; GFX11-FAKE16-NEXT:  .LBB13_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s15, s15, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s13, s13, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s15, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB13_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s15, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s14, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB13_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    s_branch .LBB13_2
+; GFX11-LABEL: bitcast_v30i32_to_v60i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s15, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v11
+; GFX11-NEXT:    s_mov_b32 s94, 0
+; GFX11-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB13_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s40, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s15, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s92, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
+; GFX11-NEXT:    s_cbranch_vccnz .LBB13_3
+; GFX11-NEXT:  .LBB13_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s14, s14, 3
+; GFX11-NEXT:    s_add_i32 s15, s15, 3
+; GFX11-NEXT:    s_add_i32 s13, s13, 3
+; GFX11-NEXT:    s_add_i32 s12, s12, 3
+; GFX11-NEXT:    s_add_i32 s11, s11, 3
+; GFX11-NEXT:    s_add_i32 s10, s10, 3
+; GFX11-NEXT:    s_add_i32 s9, s9, 3
+; GFX11-NEXT:    s_add_i32 s8, s8, 3
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s40, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s15, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s92, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s0, 16
+; GFX11-NEXT:  .LBB13_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s12, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s13, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s15, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s14, s40
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
+; GFX11-NEXT:    v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
+; GFX11-NEXT:    v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB13_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    s_branch .LBB13_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -7364,213 +7170,139 @@ define inreg <30 x i32> @bitcast_v60i16_to_v30i32_scalar(<60 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v60i16_to_v30i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v33, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v9 :: v_dual_mov_b32 v35, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v7 :: v_dual_mov_b32 v37, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v5 :: v_dual_mov_b32 v39, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v3 :: v_dual_mov_b32 v49, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v1 :: v_dual_mov_b32 v51, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v11.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v10.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v11
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB15_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB15_3
 ; GFX11-TRUE16-NEXT:  .LBB15_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v28, 16, v29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v34, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s18, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
@@ -10161,393 +9893,199 @@ define inreg <60 x half> @bitcast_v30i32_to_v60f16_scalar(<30 x i32> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr46
 ; GFX9-NEXT:    s_branch .LBB17_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v30i32_to_v60f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s94, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s15, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-TRUE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s13, s13, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s15, s15, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s15, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB17_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s15, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s14, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s40
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s15 :: v_dual_mov_b32 v19, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s13 :: v_dual_mov_b32 v21, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s11 :: v_dual_mov_b32 v23, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s9 :: v_dual_mov_b32 v25, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s7 :: v_dual_mov_b32 v27, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s6 :: v_dual_mov_b32 v29, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB17_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB17_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v30i32_to_v60f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v11
-; GFX11-FAKE16-NEXT:    s_mov_b32 s94, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB17_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s15, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB17_3
-; GFX11-FAKE16-NEXT:  .LBB17_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s15, s15, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s13, s13, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s9, s9, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s5, s5, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s29, s29, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s27, s27, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s25, s25, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s23, s23, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s21, s21, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s19, s19, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s17, s17, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s15, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB17_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s15, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s14, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB17_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    s_branch .LBB17_2
+; GFX11-LABEL: bitcast_v30i32_to_v60f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s15, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v11
+; GFX11-NEXT:    s_mov_b32 s94, 0
+; GFX11-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB17_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s40, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s15, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s92, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
+; GFX11-NEXT:    s_cbranch_vccnz .LBB17_3
+; GFX11-NEXT:  .LBB17_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s14, s14, 3
+; GFX11-NEXT:    s_add_i32 s15, s15, 3
+; GFX11-NEXT:    s_add_i32 s13, s13, 3
+; GFX11-NEXT:    s_add_i32 s12, s12, 3
+; GFX11-NEXT:    s_add_i32 s11, s11, 3
+; GFX11-NEXT:    s_add_i32 s10, s10, 3
+; GFX11-NEXT:    s_add_i32 s9, s9, 3
+; GFX11-NEXT:    s_add_i32 s8, s8, 3
+; GFX11-NEXT:    s_add_i32 s7, s7, 3
+; GFX11-NEXT:    s_add_i32 s6, s6, 3
+; GFX11-NEXT:    s_add_i32 s5, s5, 3
+; GFX11-NEXT:    s_add_i32 s4, s4, 3
+; GFX11-NEXT:    s_add_i32 s29, s29, 3
+; GFX11-NEXT:    s_add_i32 s28, s28, 3
+; GFX11-NEXT:    s_add_i32 s27, s27, 3
+; GFX11-NEXT:    s_add_i32 s26, s26, 3
+; GFX11-NEXT:    s_add_i32 s25, s25, 3
+; GFX11-NEXT:    s_add_i32 s24, s24, 3
+; GFX11-NEXT:    s_add_i32 s23, s23, 3
+; GFX11-NEXT:    s_add_i32 s22, s22, 3
+; GFX11-NEXT:    s_add_i32 s21, s21, 3
+; GFX11-NEXT:    s_add_i32 s20, s20, 3
+; GFX11-NEXT:    s_add_i32 s19, s19, 3
+; GFX11-NEXT:    s_add_i32 s18, s18, 3
+; GFX11-NEXT:    s_add_i32 s17, s17, 3
+; GFX11-NEXT:    s_add_i32 s16, s16, 3
+; GFX11-NEXT:    s_add_i32 s3, s3, 3
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s40, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s15, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s92, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s0, 16
+; GFX11-NEXT:  .LBB17_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s12, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s13, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s15, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s14, s40
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
+; GFX11-NEXT:    v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
+; GFX11-NEXT:    v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB17_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    s_branch .LBB17_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -13232,213 +12770,139 @@ define inreg <30 x i32> @bitcast_v60f16_to_v30i32_scalar(<60 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v60f16_to_v30i32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v33, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v9 :: v_dual_mov_b32 v35, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v7 :: v_dual_mov_b32 v37, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v5 :: v_dual_mov_b32 v39, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v3 :: v_dual_mov_b32 v49, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v1 :: v_dual_mov_b32 v51, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v11.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v10.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v11
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB19_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB19_3
 ; GFX11-TRUE16-NEXT:  .LBB19_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v28, 16, v29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v34, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s18 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
@@ -17263,403 +16727,204 @@ define inreg <60 x i16> @bitcast_v30f32_to_v60i16_scalar(<30 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    s_branch .LBB29_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v30f32_to_v60i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s0 :: v_dual_mov_b32 v29, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s2 :: v_dual_mov_b32 v27, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s16 :: v_dual_mov_b32 v25, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v23, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v19, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v13, s26
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s25 :: v_dual_mov_b32 v17, s28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s27 :: v_dual_mov_b32 v15, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-TRUE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v30, 1.0, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-TRUE16-NEXT:  .LBB29_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v39, v39, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v49, v49, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v81, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v80, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v71, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v70, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v38, v38, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, v49
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v48, v48, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v30
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v83, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v48
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v32 :: v_dual_mov_b32 v9, v33
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB29_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB29_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v30f32_to_v60i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s0 :: v_dual_mov_b32 v29, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s2 :: v_dual_mov_b32 v27, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s16 :: v_dual_mov_b32 v25, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v23, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v19, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v13, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s25 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB29_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB29_3
-; GFX11-FAKE16-NEXT:  .LBB29_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v30, 1.0, v30
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-FAKE16-NEXT:  .LBB29_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v39, v39, 16, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v49, v49, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v82, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v81, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v71, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v70, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v38, v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v5, v49
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v48, v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v30
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v7, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v48
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, v32 :: v_dual_mov_b32 v9, v33
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB29_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    s_branch .LBB29_2
+; GFX11-LABEL: bitcast_v30f32_to_v60i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
+; GFX11-NEXT:    v_dual_mov_b32 v30, s0 :: v_dual_mov_b32 v29, s1
+; GFX11-NEXT:    v_dual_mov_b32 v28, s2 :: v_dual_mov_b32 v27, s3
+; GFX11-NEXT:    v_dual_mov_b32 v26, s16 :: v_dual_mov_b32 v25, s17
+; GFX11-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v23, s19
+; GFX11-NEXT:    v_dual_mov_b32 v22, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT:    v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v19, s23
+; GFX11-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v13, s26
+; GFX11-NEXT:    v_dual_mov_b32 v14, s25 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB29_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB29_3
+; GFX11-NEXT:  .LBB29_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
+; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
+; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
+; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
+; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
+; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
+; GFX11-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v30, 1.0, v30
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
+; GFX11-NEXT:  .LBB29_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v35, v35, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v12, v12, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v37, v37, 16, v29
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xffff, v30
+; GFX11-NEXT:    v_lshl_or_b32 v39, v39, 16, v27
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v49, v49, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v13, v82, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v14, v81, 16, v21
+; GFX11-NEXT:    v_lshl_or_b32 v16, v71, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v70, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v38, v38, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v34, v34, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v10
+; GFX11-NEXT:    v_mov_b32_e32 v5, v49
+; GFX11-NEXT:    v_lshl_or_b32 v48, v48, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v37
+; GFX11-NEXT:    v_lshl_or_b32 v36, v36, 16, v30
+; GFX11-NEXT:    v_mov_b32_e32 v7, v31
+; GFX11-NEXT:    v_lshl_or_b32 v30, v83, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v11
+; GFX11-NEXT:    v_lshl_or_b32 v15, v80, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v36
+; GFX11-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
+; GFX11-NEXT:    v_mov_b32_e32 v4, v48
+; GFX11-NEXT:    v_mov_b32_e32 v6, v30
+; GFX11-NEXT:    v_dual_mov_b32 v8, v32 :: v_dual_mov_b32 v9, v33
+; GFX11-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB29_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr83
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr12
+; GFX11-NEXT:    ; implicit-def: $vgpr82
+; GFX11-NEXT:    ; implicit-def: $vgpr81
+; GFX11-NEXT:    ; implicit-def: $vgpr80
+; GFX11-NEXT:    ; implicit-def: $vgpr71
+; GFX11-NEXT:    ; implicit-def: $vgpr70
+; GFX11-NEXT:    ; implicit-def: $vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr68
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    s_branch .LBB29_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -19915,213 +19180,139 @@ define inreg <30 x float> @bitcast_v60i16_to_v30f32_scalar(<60 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v60i16_to_v30f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v33, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v9 :: v_dual_mov_b32 v35, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v7 :: v_dual_mov_b32 v37, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v5 :: v_dual_mov_b32 v39, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v3 :: v_dual_mov_b32 v49, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v1 :: v_dual_mov_b32 v51, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v11.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v10.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v11
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB31_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB31_3
 ; GFX11-TRUE16-NEXT:  .LBB31_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v28, 16, v29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v34, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s18, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
@@ -22684,403 +21875,204 @@ define inreg <60 x half> @bitcast_v30f32_to_v60f16_scalar(<30 x float> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    s_branch .LBB33_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v30f32_to_v60f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s0 :: v_dual_mov_b32 v29, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s2 :: v_dual_mov_b32 v27, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s16 :: v_dual_mov_b32 v25, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v23, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v19, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v13, s26
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s25 :: v_dual_mov_b32 v17, s28
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s27 :: v_dual_mov_b32 v15, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-TRUE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v16, 1.0, v16
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-TRUE16-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v30, 1.0, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-TRUE16-NEXT:  .LBB33_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v39, v39, 16, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v49, v49, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v81, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v80, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v71, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v70, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v38, v38, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, v49
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v48, v48, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v30
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v83, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v36
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v48
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v32 :: v_dual_mov_b32 v9, v33
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB33_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB33_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v30f32_to_v60f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s0 :: v_dual_mov_b32 v29, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s2 :: v_dual_mov_b32 v27, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s16 :: v_dual_mov_b32 v25, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v23, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s20 :: v_dual_mov_b32 v21, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v19, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v13, s26
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s25 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB33_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB33_3
-; GFX11-FAKE16-NEXT:  .LBB33_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v18, 1.0, v18
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
-; GFX11-FAKE16-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v30, 1.0, v30
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-FAKE16-NEXT:  .LBB33_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v39, v39, 16, v27
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v49, v49, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v82, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v81, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v71, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v70, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v38, v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v5, v49
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v48, v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v30
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v7, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v48
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, v32 :: v_dual_mov_b32 v9, v33
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB33_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr31
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    s_branch .LBB33_2
+; GFX11-LABEL: bitcast_v30f32_to_v60f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
+; GFX11-NEXT:    v_dual_mov_b32 v30, s0 :: v_dual_mov_b32 v29, s1
+; GFX11-NEXT:    v_dual_mov_b32 v28, s2 :: v_dual_mov_b32 v27, s3
+; GFX11-NEXT:    v_dual_mov_b32 v26, s16 :: v_dual_mov_b32 v25, s17
+; GFX11-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v23, s19
+; GFX11-NEXT:    v_dual_mov_b32 v22, s20 :: v_dual_mov_b32 v21, s21
+; GFX11-NEXT:    v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v19, s23
+; GFX11-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v13, s26
+; GFX11-NEXT:    v_dual_mov_b32 v14, s25 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v17, s28 :: v_dual_mov_b32 v16, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB33_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB33_3
+; GFX11-NEXT:  .LBB33_2: ; %cmp.true
+; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
+; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
+; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
+; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
+; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
+; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
+; GFX11-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v17, 1.0, v17
+; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
+; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v18, 1.0, v18
+; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v20, 1.0, v20
+; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v22, 1.0, v22
+; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v24, 1.0, v24
+; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v26, 1.0, v26
+; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v28, 1.0, v28
+; GFX11-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v30, 1.0, v30
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
+; GFX11-NEXT:  .LBB33_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-NEXT:    v_and_b32_e32 v27, 0xffff, v27
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v13
+; GFX11-NEXT:    v_lshl_or_b32 v35, v35, 16, v19
+; GFX11-NEXT:    v_lshl_or_b32 v12, v12, 16, v18
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v37, v37, 16, v29
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xffff, v30
+; GFX11-NEXT:    v_lshl_or_b32 v39, v39, 16, v27
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v49, v49, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-NEXT:    v_lshl_or_b32 v31, v31, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_lshl_or_b32 v13, v82, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v14, v81, 16, v21
+; GFX11-NEXT:    v_lshl_or_b32 v16, v71, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v17, v70, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v38, v38, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v22
+; GFX11-NEXT:    v_lshl_or_b32 v34, v34, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v10
+; GFX11-NEXT:    v_mov_b32_e32 v5, v49
+; GFX11-NEXT:    v_lshl_or_b32 v48, v48, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v37
+; GFX11-NEXT:    v_lshl_or_b32 v36, v36, 16, v30
+; GFX11-NEXT:    v_mov_b32_e32 v7, v31
+; GFX11-NEXT:    v_lshl_or_b32 v30, v83, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v11
+; GFX11-NEXT:    v_lshl_or_b32 v15, v80, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v36
+; GFX11-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
+; GFX11-NEXT:    v_mov_b32_e32 v4, v48
+; GFX11-NEXT:    v_mov_b32_e32 v6, v30
+; GFX11-NEXT:    v_dual_mov_b32 v8, v32 :: v_dual_mov_b32 v9, v33
+; GFX11-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB33_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr83
+; GFX11-NEXT:    ; implicit-def: $vgpr31
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr12
+; GFX11-NEXT:    ; implicit-def: $vgpr82
+; GFX11-NEXT:    ; implicit-def: $vgpr81
+; GFX11-NEXT:    ; implicit-def: $vgpr80
+; GFX11-NEXT:    ; implicit-def: $vgpr71
+; GFX11-NEXT:    ; implicit-def: $vgpr70
+; GFX11-NEXT:    ; implicit-def: $vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr68
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    s_branch .LBB33_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -25765,213 +24757,139 @@ define inreg <30 x float> @bitcast_v60f16_to_v30f32_scalar(<60 x half> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v60f16_to_v30f32_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v33, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v9 :: v_dual_mov_b32 v35, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v7 :: v_dual_mov_b32 v37, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v5 :: v_dual_mov_b32 v39, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v3 :: v_dual_mov_b32 v49, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v1 :: v_dual_mov_b32 v51, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v11.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v10.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v11
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB35_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB35_3
 ; GFX11-TRUE16-NEXT:  .LBB35_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v28, 16, v29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v34, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s18 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
@@ -28926,393 +27844,199 @@ define inreg <60 x i16> @bitcast_v15i64_to_v60i16_scalar(<15 x i64> inreg %a, i3
 ; GFX9-NEXT:    ; implicit-def: $sgpr46
 ; GFX9-NEXT:    s_branch .LBB41_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v15i64_to_v60i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s94, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s15, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-TRUE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s8, s8, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s10, s10, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s13, s13, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s12, s12, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s15, s15, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s14, s14, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s15, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB41_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s15, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s14, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s40
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s15 :: v_dual_mov_b32 v19, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s13 :: v_dual_mov_b32 v21, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s11 :: v_dual_mov_b32 v23, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s9 :: v_dual_mov_b32 v25, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s7 :: v_dual_mov_b32 v27, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s6 :: v_dual_mov_b32 v29, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB41_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB41_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v15i64_to_v60i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v11
-; GFX11-FAKE16-NEXT:    s_mov_b32 s94, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB41_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s15, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB41_3
-; GFX11-FAKE16-NEXT:  .LBB41_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s15, s15, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s14, s14, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s12, s12, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s13, s13, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s11, s11, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s9, s9, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s7, s7, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s15, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB41_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s15, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s14, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB41_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    s_branch .LBB41_2
+; GFX11-LABEL: bitcast_v15i64_to_v60i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s15, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v11
+; GFX11-NEXT:    s_mov_b32 s94, 0
+; GFX11-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB41_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s40, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s15, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s92, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
+; GFX11-NEXT:    s_cbranch_vccnz .LBB41_3
+; GFX11-NEXT:  .LBB41_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s15, s15, 3
+; GFX11-NEXT:    s_addc_u32 s14, s14, 0
+; GFX11-NEXT:    s_add_u32 s12, s12, 3
+; GFX11-NEXT:    s_addc_u32 s13, s13, 0
+; GFX11-NEXT:    s_add_u32 s10, s10, 3
+; GFX11-NEXT:    s_addc_u32 s11, s11, 0
+; GFX11-NEXT:    s_add_u32 s8, s8, 3
+; GFX11-NEXT:    s_addc_u32 s9, s9, 0
+; GFX11-NEXT:    s_add_u32 s6, s6, 3
+; GFX11-NEXT:    s_addc_u32 s7, s7, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s40, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s15, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s92, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s0, 16
+; GFX11-NEXT:  .LBB41_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s12, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s13, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s15, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s14, s40
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
+; GFX11-NEXT:    v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
+; GFX11-NEXT:    v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB41_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    s_branch .LBB41_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -31568,213 +30292,139 @@ define inreg <15 x i64> @bitcast_v60i16_to_v15i64_scalar(<60 x i16> inreg %a, i3
 ; GFX11-TRUE16-LABEL: bitcast_v60i16_to_v15i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v33, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v9 :: v_dual_mov_b32 v35, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v7 :: v_dual_mov_b32 v37, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v5 :: v_dual_mov_b32 v39, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v3 :: v_dual_mov_b32 v49, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v1 :: v_dual_mov_b32 v51, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v11.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v10.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v11
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB43_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB43_3
 ; GFX11-TRUE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v28, 16, v29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v34, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s18, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
@@ -34382,393 +33032,199 @@ define inreg <60 x half> @bitcast_v15i64_to_v60f16_scalar(<15 x i64> inreg %a, i
 ; GFX9-NEXT:    ; implicit-def: $sgpr46
 ; GFX9-NEXT:    s_branch .LBB45_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v15i64_to_v60f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v1
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v2
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v3
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v4
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v5
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v6
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v7
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v9
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v10
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s94, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s15, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-TRUE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s4, s4, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s7, s7, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s9, s9, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s8, s8, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s11, s11, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s10, s10, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s13, s13, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s12, s12, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s15, s15, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s14, s14, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s4, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s6, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s5, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s9, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s10, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s47, s11, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s56, s12, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s57, s13, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s58, s14, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s59, s15, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-TRUE16-NEXT:  .LBB45_3: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s15, s59
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s14, s58
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s57
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s56
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s47
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s46
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s40
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s15 :: v_dual_mov_b32 v19, s14
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s13 :: v_dual_mov_b32 v21, s12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s11 :: v_dual_mov_b32 v23, s10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s9 :: v_dual_mov_b32 v25, s8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s7 :: v_dual_mov_b32 v27, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s6 :: v_dual_mov_b32 v29, s4
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB45_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr93_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr91_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr89_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr79_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr77_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr75_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr73_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr63_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr61_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr59_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr57_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr47_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr45_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr41_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr40_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB45_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v15i64_to_v60f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v3
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v4
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v5
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v6
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v7
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v8
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v9
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v10
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v11
-; GFX11-FAKE16-NEXT:    s_mov_b32 s94, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB45_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s15, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB45_3
-; GFX11-FAKE16-NEXT:  .LBB45_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_u32 s15, s15, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s14, s14, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s12, s12, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s13, s13, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s10, s10, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s11, s11, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s8, s8, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s9, s9, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s6, s6, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s7, s7, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s4, s4, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s5, s5, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s28, s28, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s29, s29, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s26, s26, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s27, s27, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s24, s24, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s25, s25, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s22, s22, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s23, s23, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s20, s20, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s21, s21, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s18, s18, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s19, s19, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s16, s16, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s17, s17, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, 0
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s40, s14, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s41, s15, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s13, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s43, s12, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s44, s11, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s45, s10, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s46, s9, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s47, s8, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s56, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s57, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s58, s5, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s59, s4, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s60, s29, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s61, s28, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s62, s27, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s63, s26, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s72, s25, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s73, s24, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s74, s23, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s75, s22, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s76, s21, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s77, s20, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s78, s19, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s79, s18, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s88, s17, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s89, s16, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s90, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s91, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s92, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s93, s0, 16
-; GFX11-FAKE16-NEXT:  .LBB45_3: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s59
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s58
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s6, s6, s57
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s7, s7, s56
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s8, s8, s47
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s9, s9, s46
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s10, s10, s45
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s11, s11, s44
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s12, s12, s43
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s13, s13, s42
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s15, s15, s41
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s14, s14, s40
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB45_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr93
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr91
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr89
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr79
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr77
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr75
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr73
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr63
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr61
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr59
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr57
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr47
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr45
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr41
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr40
-; GFX11-FAKE16-NEXT:    s_branch .LBB45_2
+; GFX11-LABEL: bitcast_v15i64_to_v60f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s8, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s9, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s10, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s11, v7
+; GFX11-NEXT:    v_readfirstlane_b32 s12, v8
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v9
+; GFX11-NEXT:    v_readfirstlane_b32 s15, v10
+; GFX11-NEXT:    v_readfirstlane_b32 s14, v11
+; GFX11-NEXT:    s_mov_b32 s94, 0
+; GFX11-NEXT:    s_and_b32 s40, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB45_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s40, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s15, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s92, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s0, 16
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s94
+; GFX11-NEXT:    s_cbranch_vccnz .LBB45_3
+; GFX11-NEXT:  .LBB45_2: ; %cmp.true
+; GFX11-NEXT:    s_add_u32 s15, s15, 3
+; GFX11-NEXT:    s_addc_u32 s14, s14, 0
+; GFX11-NEXT:    s_add_u32 s12, s12, 3
+; GFX11-NEXT:    s_addc_u32 s13, s13, 0
+; GFX11-NEXT:    s_add_u32 s10, s10, 3
+; GFX11-NEXT:    s_addc_u32 s11, s11, 0
+; GFX11-NEXT:    s_add_u32 s8, s8, 3
+; GFX11-NEXT:    s_addc_u32 s9, s9, 0
+; GFX11-NEXT:    s_add_u32 s6, s6, 3
+; GFX11-NEXT:    s_addc_u32 s7, s7, 0
+; GFX11-NEXT:    s_add_u32 s4, s4, 3
+; GFX11-NEXT:    s_addc_u32 s5, s5, 0
+; GFX11-NEXT:    s_add_u32 s28, s28, 3
+; GFX11-NEXT:    s_addc_u32 s29, s29, 0
+; GFX11-NEXT:    s_add_u32 s26, s26, 3
+; GFX11-NEXT:    s_addc_u32 s27, s27, 0
+; GFX11-NEXT:    s_add_u32 s24, s24, 3
+; GFX11-NEXT:    s_addc_u32 s25, s25, 0
+; GFX11-NEXT:    s_add_u32 s22, s22, 3
+; GFX11-NEXT:    s_addc_u32 s23, s23, 0
+; GFX11-NEXT:    s_add_u32 s20, s20, 3
+; GFX11-NEXT:    s_addc_u32 s21, s21, 0
+; GFX11-NEXT:    s_add_u32 s18, s18, 3
+; GFX11-NEXT:    s_addc_u32 s19, s19, 0
+; GFX11-NEXT:    s_add_u32 s16, s16, 3
+; GFX11-NEXT:    s_addc_u32 s17, s17, 0
+; GFX11-NEXT:    s_add_u32 s2, s2, 3
+; GFX11-NEXT:    s_addc_u32 s3, s3, 0
+; GFX11-NEXT:    s_add_u32 s0, s0, 3
+; GFX11-NEXT:    s_addc_u32 s1, s1, 0
+; GFX11-NEXT:    s_lshr_b32 s40, s14, 16
+; GFX11-NEXT:    s_lshr_b32 s41, s15, 16
+; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
+; GFX11-NEXT:    s_lshr_b32 s43, s12, 16
+; GFX11-NEXT:    s_lshr_b32 s44, s11, 16
+; GFX11-NEXT:    s_lshr_b32 s45, s10, 16
+; GFX11-NEXT:    s_lshr_b32 s46, s9, 16
+; GFX11-NEXT:    s_lshr_b32 s47, s8, 16
+; GFX11-NEXT:    s_lshr_b32 s56, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s57, s6, 16
+; GFX11-NEXT:    s_lshr_b32 s58, s5, 16
+; GFX11-NEXT:    s_lshr_b32 s59, s4, 16
+; GFX11-NEXT:    s_lshr_b32 s60, s29, 16
+; GFX11-NEXT:    s_lshr_b32 s61, s28, 16
+; GFX11-NEXT:    s_lshr_b32 s62, s27, 16
+; GFX11-NEXT:    s_lshr_b32 s63, s26, 16
+; GFX11-NEXT:    s_lshr_b32 s72, s25, 16
+; GFX11-NEXT:    s_lshr_b32 s73, s24, 16
+; GFX11-NEXT:    s_lshr_b32 s74, s23, 16
+; GFX11-NEXT:    s_lshr_b32 s75, s22, 16
+; GFX11-NEXT:    s_lshr_b32 s76, s21, 16
+; GFX11-NEXT:    s_lshr_b32 s77, s20, 16
+; GFX11-NEXT:    s_lshr_b32 s78, s19, 16
+; GFX11-NEXT:    s_lshr_b32 s79, s18, 16
+; GFX11-NEXT:    s_lshr_b32 s88, s17, 16
+; GFX11-NEXT:    s_lshr_b32 s89, s16, 16
+; GFX11-NEXT:    s_lshr_b32 s90, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s91, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s92, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s93, s0, 16
+; GFX11-NEXT:  .LBB45_3: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s0, s93
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s1, s92
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s91
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s3, s90
+; GFX11-NEXT:    s_pack_ll_b32_b16 s16, s16, s89
+; GFX11-NEXT:    s_pack_ll_b32_b16 s17, s17, s88
+; GFX11-NEXT:    s_pack_ll_b32_b16 s18, s18, s79
+; GFX11-NEXT:    s_pack_ll_b32_b16 s19, s19, s78
+; GFX11-NEXT:    s_pack_ll_b32_b16 s20, s20, s77
+; GFX11-NEXT:    s_pack_ll_b32_b16 s21, s21, s76
+; GFX11-NEXT:    s_pack_ll_b32_b16 s22, s22, s75
+; GFX11-NEXT:    s_pack_ll_b32_b16 s23, s23, s74
+; GFX11-NEXT:    s_pack_ll_b32_b16 s24, s24, s73
+; GFX11-NEXT:    s_pack_ll_b32_b16 s25, s25, s72
+; GFX11-NEXT:    s_pack_ll_b32_b16 s26, s26, s63
+; GFX11-NEXT:    s_pack_ll_b32_b16 s27, s27, s62
+; GFX11-NEXT:    s_pack_ll_b32_b16 s28, s28, s61
+; GFX11-NEXT:    s_pack_ll_b32_b16 s29, s29, s60
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s4, s59
+; GFX11-NEXT:    s_pack_ll_b32_b16 s5, s5, s58
+; GFX11-NEXT:    s_pack_ll_b32_b16 s6, s6, s57
+; GFX11-NEXT:    s_pack_ll_b32_b16 s7, s7, s56
+; GFX11-NEXT:    s_pack_ll_b32_b16 s8, s8, s47
+; GFX11-NEXT:    s_pack_ll_b32_b16 s9, s9, s46
+; GFX11-NEXT:    s_pack_ll_b32_b16 s10, s10, s45
+; GFX11-NEXT:    s_pack_ll_b32_b16 s11, s11, s44
+; GFX11-NEXT:    s_pack_ll_b32_b16 s12, s12, s43
+; GFX11-NEXT:    s_pack_ll_b32_b16 s13, s13, s42
+; GFX11-NEXT:    s_pack_ll_b32_b16 s15, s15, s41
+; GFX11-NEXT:    s_pack_ll_b32_b16 s14, s14, s40
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
+; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
+; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
+; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    v_dual_mov_b32 v18, s4 :: v_dual_mov_b32 v19, s5
+; GFX11-NEXT:    v_dual_mov_b32 v20, s6 :: v_dual_mov_b32 v21, s7
+; GFX11-NEXT:    v_dual_mov_b32 v22, s8 :: v_dual_mov_b32 v23, s9
+; GFX11-NEXT:    v_dual_mov_b32 v24, s10 :: v_dual_mov_b32 v25, s11
+; GFX11-NEXT:    v_dual_mov_b32 v26, s12 :: v_dual_mov_b32 v27, s13
+; GFX11-NEXT:    v_dual_mov_b32 v28, s15 :: v_dual_mov_b32 v29, s14
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB45_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr93
+; GFX11-NEXT:    ; implicit-def: $sgpr92
+; GFX11-NEXT:    ; implicit-def: $sgpr91
+; GFX11-NEXT:    ; implicit-def: $sgpr90
+; GFX11-NEXT:    ; implicit-def: $sgpr89
+; GFX11-NEXT:    ; implicit-def: $sgpr88
+; GFX11-NEXT:    ; implicit-def: $sgpr79
+; GFX11-NEXT:    ; implicit-def: $sgpr78
+; GFX11-NEXT:    ; implicit-def: $sgpr77
+; GFX11-NEXT:    ; implicit-def: $sgpr76
+; GFX11-NEXT:    ; implicit-def: $sgpr75
+; GFX11-NEXT:    ; implicit-def: $sgpr74
+; GFX11-NEXT:    ; implicit-def: $sgpr73
+; GFX11-NEXT:    ; implicit-def: $sgpr72
+; GFX11-NEXT:    ; implicit-def: $sgpr63
+; GFX11-NEXT:    ; implicit-def: $sgpr62
+; GFX11-NEXT:    ; implicit-def: $sgpr61
+; GFX11-NEXT:    ; implicit-def: $sgpr60
+; GFX11-NEXT:    ; implicit-def: $sgpr59
+; GFX11-NEXT:    ; implicit-def: $sgpr58
+; GFX11-NEXT:    ; implicit-def: $sgpr57
+; GFX11-NEXT:    ; implicit-def: $sgpr56
+; GFX11-NEXT:    ; implicit-def: $sgpr47
+; GFX11-NEXT:    ; implicit-def: $sgpr46
+; GFX11-NEXT:    ; implicit-def: $sgpr45
+; GFX11-NEXT:    ; implicit-def: $sgpr44
+; GFX11-NEXT:    ; implicit-def: $sgpr43
+; GFX11-NEXT:    ; implicit-def: $sgpr42
+; GFX11-NEXT:    ; implicit-def: $sgpr41
+; GFX11-NEXT:    ; implicit-def: $sgpr40
+; GFX11-NEXT:    s_branch .LBB45_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -37453,213 +35909,139 @@ define inreg <15 x i64> @bitcast_v60f16_to_v15i64_scalar(<60 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v60f16_to_v15i64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v33, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v9 :: v_dual_mov_b32 v35, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v7 :: v_dual_mov_b32 v37, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v5 :: v_dual_mov_b32 v39, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v3 :: v_dual_mov_b32 v49, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v1 :: v_dual_mov_b32 v51, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v11.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v10.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v11
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB47_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB47_3
 ; GFX11-TRUE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v28, 16, v29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v34, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s18 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
@@ -39575,400 +37957,204 @@ define inreg <60 x i16> @bitcast_v15f64_to_v60i16_scalar(<15 x double> inreg %a,
 ; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    s_branch .LBB49_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v15f64_to_v60i16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v29, s0 :: v_dual_mov_b32 v30, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v27, s2 :: v_dual_mov_b32 v28, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v25, s16 :: v_dual_mov_b32 v26, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v23, s18 :: v_dual_mov_b32 v24, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v21, s20 :: v_dual_mov_b32 v22, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v19, s22 :: v_dual_mov_b32 v20, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s24 :: v_dual_mov_b32 v14, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v18, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v29
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-TRUE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[29:30], v[29:30], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v29
-; GFX11-TRUE16-NEXT:  .LBB49_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v81, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v80, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v70, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v48, v48, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v71, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v32 :: v_dual_and_b32 v1, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v83, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v49, v49, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, v49
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v38, v38, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v39, v39, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_and_b32 v4, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v36 :: v_dual_mov_b32 v1, v37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v48 :: v_dual_mov_b32 v9, v33
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB49_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB49_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v15f64_to_v60i16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s0 :: v_dual_mov_b32 v31, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s2 :: v_dual_mov_b32 v29, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s16 :: v_dual_mov_b32 v27, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v25, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s20 :: v_dual_mov_b32 v23, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v21, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v19, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB49_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB49_3
-; GFX11-FAKE16-NEXT:  .LBB49_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-FAKE16-NEXT:  .LBB49_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v48, v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v31
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v39, v39, 16, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v82, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v38, v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v7, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v49, v49, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v30
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v9, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v81, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v71, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v70, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, v48 :: v_dual_mov_b32 v5, v49
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v32
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB49_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    s_branch .LBB49_2
+; GFX11-LABEL: bitcast_v15f64_to_v60i16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
+; GFX11-NEXT:    v_dual_mov_b32 v30, s0 :: v_dual_mov_b32 v31, s1
+; GFX11-NEXT:    v_dual_mov_b32 v28, s2 :: v_dual_mov_b32 v29, s3
+; GFX11-NEXT:    v_dual_mov_b32 v26, s16 :: v_dual_mov_b32 v27, s17
+; GFX11-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v25, s19
+; GFX11-NEXT:    v_dual_mov_b32 v22, s20 :: v_dual_mov_b32 v23, s21
+; GFX11-NEXT:    v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v21, s23
+; GFX11-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v19, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB49_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB49_3
+; GFX11-NEXT:  .LBB49_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
+; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
+; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
+; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
+; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
+; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
+; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
+; GFX11-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
+; GFX11-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
+; GFX11-NEXT:  .LBB49_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-NEXT:    v_lshl_or_b32 v34, v34, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_lshl_or_b32 v48, v48, 16, v26
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_lshl_or_b32 v13, v13, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v37, v37, 16, v31
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xffff, v30
+; GFX11-NEXT:    v_lshl_or_b32 v39, v39, 16, v29
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_lshl_or_b32 v31, v82, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v27
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v35, v35, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v12, v12, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v38, v38, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v22
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v10
+; GFX11-NEXT:    v_mov_b32_e32 v7, v31
+; GFX11-NEXT:    v_lshl_or_b32 v49, v49, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v37
+; GFX11-NEXT:    v_lshl_or_b32 v36, v36, 16, v30
+; GFX11-NEXT:    v_mov_b32_e32 v9, v33
+; GFX11-NEXT:    v_lshl_or_b32 v30, v83, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v11
+; GFX11-NEXT:    v_lshl_or_b32 v14, v81, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v15, v80, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v71, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v70, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v36
+; GFX11-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
+; GFX11-NEXT:    v_dual_mov_b32 v4, v48 :: v_dual_mov_b32 v5, v49
+; GFX11-NEXT:    v_mov_b32_e32 v6, v30
+; GFX11-NEXT:    v_mov_b32_e32 v8, v32
+; GFX11-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB49_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr83
+; GFX11-NEXT:    ; implicit-def: $vgpr82
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr12
+; GFX11-NEXT:    ; implicit-def: $vgpr13
+; GFX11-NEXT:    ; implicit-def: $vgpr81
+; GFX11-NEXT:    ; implicit-def: $vgpr80
+; GFX11-NEXT:    ; implicit-def: $vgpr71
+; GFX11-NEXT:    ; implicit-def: $vgpr70
+; GFX11-NEXT:    ; implicit-def: $vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr68
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    s_branch .LBB49_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -42224,213 +40410,139 @@ define inreg <15 x double> @bitcast_v60i16_to_v15f64_scalar(<60 x i16> inreg %a,
 ; GFX11-TRUE16-LABEL: bitcast_v60i16_to_v15f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v33, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v9 :: v_dual_mov_b32 v35, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v7 :: v_dual_mov_b32 v37, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v5 :: v_dual_mov_b32 v39, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v3 :: v_dual_mov_b32 v49, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v1 :: v_dual_mov_b32 v51, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v11.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v10.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v11
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB51_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB51_3
 ; GFX11-TRUE16-NEXT:  .LBB51_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v20, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v28, 16, v29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v34, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s4, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, s5, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, s7, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s17, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s18, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
@@ -44900,400 +43012,204 @@ define inreg <60 x half> @bitcast_v15f64_to_v60f16_scalar(<15 x double> inreg %a
 ; GFX9-NEXT:    ; implicit-def: $vgpr54
 ; GFX9-NEXT:    s_branch .LBB53_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v15f64_to_v60f16_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v29, s0 :: v_dual_mov_b32 v30, s1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v27, s2 :: v_dual_mov_b32 v28, s3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v25, s16 :: v_dual_mov_b32 v26, s17
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v23, s18 :: v_dual_mov_b32 v24, s19
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v21, s20 :: v_dual_mov_b32 v22, s21
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v19, s22 :: v_dual_mov_b32 v20, s23
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s24 :: v_dual_mov_b32 v14, s25
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v18, s27
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v29
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-TRUE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
-; GFX11-TRUE16-NEXT:    v_add_f64 v[29:30], v[29:30], 1.0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v15
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v29
-; GFX11-TRUE16-NEXT:  .LBB53_3: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v19
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v13
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v82, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v81, 16, v17
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v21
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v80, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v70, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v48, v48, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v71, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v32 :: v_dual_and_b32 v1, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v83, 16, v23
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v49, v49, 16, v25
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, v49
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v29
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v38, v38, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v39, v39, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_and_b32 v4, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v36 :: v_dual_mov_b32 v1, v37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v48 :: v_dual_mov_b32 v9, v33
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB53_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB53_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v15f64_to_v60f16_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s0 :: v_dual_mov_b32 v31, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s2 :: v_dual_mov_b32 v29, s3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s16 :: v_dual_mov_b32 v27, s17
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v25, s19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s20 :: v_dual_mov_b32 v23, s21
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v21, s23
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v19, s25
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB53_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB53_3
-; GFX11-FAKE16-NEXT:  .LBB53_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
-; GFX11-FAKE16-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v21
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v25
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v27
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v29
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v31
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
-; GFX11-FAKE16-NEXT:  .LBB53_3: ; %end
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v34, v34, 16, v20
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v48, v48, 16, v26
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v13, 16, v19
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v37, v37, 16, v31
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v39, v39, 16, v29
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v82, 16, v25
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v27
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v33, v33, 16, v23
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v35, v35, 16, v21
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v2
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v5
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v6
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v38, v38, 16, v28
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v32, 16, v22
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v9
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v10
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v7, v31
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v49, v49, 16, v26
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, v37
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v36, v36, 16, v30
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v9, v33
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v83, 16, v24
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v11
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v81, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v80, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v71, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v70, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v36
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, v48 :: v_dual_mov_b32 v5, v49
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v32
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB53_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
-; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
-; GFX11-FAKE16-NEXT:    s_branch .LBB53_2
+; GFX11-LABEL: bitcast_v15f64_to_v60f16_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
+; GFX11-NEXT:    v_dual_mov_b32 v30, s0 :: v_dual_mov_b32 v31, s1
+; GFX11-NEXT:    v_dual_mov_b32 v28, s2 :: v_dual_mov_b32 v29, s3
+; GFX11-NEXT:    v_dual_mov_b32 v26, s16 :: v_dual_mov_b32 v27, s17
+; GFX11-NEXT:    v_dual_mov_b32 v24, s18 :: v_dual_mov_b32 v25, s19
+; GFX11-NEXT:    v_dual_mov_b32 v22, s20 :: v_dual_mov_b32 v23, s21
+; GFX11-NEXT:    v_dual_mov_b32 v20, s22 :: v_dual_mov_b32 v21, s23
+; GFX11-NEXT:    v_dual_mov_b32 v18, s24 :: v_dual_mov_b32 v19, s25
+; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
+; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_and_b32 s1, vcc_lo, exec_lo
+; GFX11-NEXT:    s_cbranch_scc0 .LBB53_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB53_3
+; GFX11-NEXT:  .LBB53_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
+; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
+; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
+; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
+; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
+; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
+; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
+; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
+; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
+; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
+; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
+; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
+; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
+; GFX11-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
+; GFX11-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v51, 16, v10
+; GFX11-NEXT:    v_lshrrev_b32_e32 v52, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v53, 16, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
+; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v6
+; GFX11-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v65, 16, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v66, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v67, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v68, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v12, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 16, v21
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 16, v20
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 16, v22
+; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 16, v25
+; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v24
+; GFX11-NEXT:    v_lshrrev_b32_e32 v49, 16, v27
+; GFX11-NEXT:    v_lshrrev_b32_e32 v48, 16, v26
+; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v29
+; GFX11-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 16, v31
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
+; GFX11-NEXT:  .LBB53_3: ; %end
+; GFX11-NEXT:    v_and_b32_e32 v20, 0xffff, v20
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v26
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-NEXT:    v_and_b32_e32 v31, 0xffff, v31
+; GFX11-NEXT:    v_and_b32_e32 v29, 0xffff, v29
+; GFX11-NEXT:    v_lshl_or_b32 v34, v34, 16, v20
+; GFX11-NEXT:    v_and_b32_e32 v19, 0xffff, v19
+; GFX11-NEXT:    v_lshl_or_b32 v48, v48, 16, v26
+; GFX11-NEXT:    v_and_b32_e32 v25, 0xffff, v25
+; GFX11-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-NEXT:    v_and_b32_e32 v21, 0xffff, v21
+; GFX11-NEXT:    v_and_b32_e32 v18, 0xffff, v18
+; GFX11-NEXT:    v_lshl_or_b32 v13, v13, 16, v19
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-NEXT:    v_lshl_or_b32 v19, v68, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v3
+; GFX11-NEXT:    v_lshl_or_b32 v37, v37, 16, v31
+; GFX11-NEXT:    v_and_b32_e32 v30, 0xffff, v30
+; GFX11-NEXT:    v_lshl_or_b32 v39, v39, 16, v29
+; GFX11-NEXT:    v_and_b32_e32 v28, 0xffff, v28
+; GFX11-NEXT:    v_lshl_or_b32 v31, v82, 16, v25
+; GFX11-NEXT:    v_and_b32_e32 v26, 0xffff, v27
+; GFX11-NEXT:    v_lshl_or_b32 v33, v33, 16, v23
+; GFX11-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-NEXT:    v_lshl_or_b32 v35, v35, 16, v21
+; GFX11-NEXT:    v_and_b32_e32 v22, 0xffff, v22
+; GFX11-NEXT:    v_lshl_or_b32 v12, v12, 16, v18
+; GFX11-NEXT:    v_lshl_or_b32 v18, v69, 16, v0
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v2
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v4
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v5
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v6
+; GFX11-NEXT:    v_lshl_or_b32 v21, v66, 16, v1
+; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v8
+; GFX11-NEXT:    v_lshl_or_b32 v38, v38, 16, v28
+; GFX11-NEXT:    v_lshl_or_b32 v32, v32, 16, v22
+; GFX11-NEXT:    v_and_b32_e32 v14, 0xffff, v14
+; GFX11-NEXT:    v_and_b32_e32 v15, 0xffff, v15
+; GFX11-NEXT:    v_and_b32_e32 v16, 0xffff, v16
+; GFX11-NEXT:    v_and_b32_e32 v17, 0xffff, v17
+; GFX11-NEXT:    v_lshl_or_b32 v20, v67, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v22, v65, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v23, v64, 16, v3
+; GFX11-NEXT:    v_and_b32_e32 v0, 0xffff, v7
+; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v9
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v10
+; GFX11-NEXT:    v_mov_b32_e32 v7, v31
+; GFX11-NEXT:    v_lshl_or_b32 v49, v49, 16, v26
+; GFX11-NEXT:    v_lshl_or_b32 v26, v53, 16, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v37
+; GFX11-NEXT:    v_lshl_or_b32 v36, v36, 16, v30
+; GFX11-NEXT:    v_mov_b32_e32 v9, v33
+; GFX11-NEXT:    v_lshl_or_b32 v30, v83, 16, v24
+; GFX11-NEXT:    v_lshl_or_b32 v24, v55, 16, v4
+; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v11
+; GFX11-NEXT:    v_lshl_or_b32 v14, v81, 16, v14
+; GFX11-NEXT:    v_lshl_or_b32 v15, v80, 16, v15
+; GFX11-NEXT:    v_lshl_or_b32 v16, v71, 16, v16
+; GFX11-NEXT:    v_lshl_or_b32 v17, v70, 16, v17
+; GFX11-NEXT:    v_lshl_or_b32 v25, v54, 16, v0
+; GFX11-NEXT:    v_lshl_or_b32 v27, v52, 16, v2
+; GFX11-NEXT:    v_lshl_or_b32 v28, v51, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v29, v50, 16, v4
+; GFX11-NEXT:    v_mov_b32_e32 v0, v36
+; GFX11-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
+; GFX11-NEXT:    v_dual_mov_b32 v4, v48 :: v_dual_mov_b32 v5, v49
+; GFX11-NEXT:    v_mov_b32_e32 v6, v30
+; GFX11-NEXT:    v_mov_b32_e32 v8, v32
+; GFX11-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB53_4:
+; GFX11-NEXT:    ; implicit-def: $vgpr36
+; GFX11-NEXT:    ; implicit-def: $vgpr37
+; GFX11-NEXT:    ; implicit-def: $vgpr38
+; GFX11-NEXT:    ; implicit-def: $vgpr39
+; GFX11-NEXT:    ; implicit-def: $vgpr48
+; GFX11-NEXT:    ; implicit-def: $vgpr49
+; GFX11-NEXT:    ; implicit-def: $vgpr83
+; GFX11-NEXT:    ; implicit-def: $vgpr82
+; GFX11-NEXT:    ; implicit-def: $vgpr32
+; GFX11-NEXT:    ; implicit-def: $vgpr33
+; GFX11-NEXT:    ; implicit-def: $vgpr34
+; GFX11-NEXT:    ; implicit-def: $vgpr35
+; GFX11-NEXT:    ; implicit-def: $vgpr12
+; GFX11-NEXT:    ; implicit-def: $vgpr13
+; GFX11-NEXT:    ; implicit-def: $vgpr81
+; GFX11-NEXT:    ; implicit-def: $vgpr80
+; GFX11-NEXT:    ; implicit-def: $vgpr71
+; GFX11-NEXT:    ; implicit-def: $vgpr70
+; GFX11-NEXT:    ; implicit-def: $vgpr69
+; GFX11-NEXT:    ; implicit-def: $vgpr68
+; GFX11-NEXT:    ; implicit-def: $vgpr67
+; GFX11-NEXT:    ; implicit-def: $vgpr66
+; GFX11-NEXT:    ; implicit-def: $vgpr65
+; GFX11-NEXT:    ; implicit-def: $vgpr64
+; GFX11-NEXT:    ; implicit-def: $vgpr55
+; GFX11-NEXT:    ; implicit-def: $vgpr54
+; GFX11-NEXT:    ; implicit-def: $vgpr53
+; GFX11-NEXT:    ; implicit-def: $vgpr52
+; GFX11-NEXT:    ; implicit-def: $vgpr51
+; GFX11-NEXT:    ; implicit-def: $vgpr50
+; GFX11-NEXT:    s_branch .LBB53_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -47978,213 +45894,139 @@ define inreg <15 x double> @bitcast_v60f16_to_v15f64_scalar(<60 x half> inreg %a
 ; GFX11-TRUE16-LABEL: bitcast_v60f16_to_v15f64_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v33, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, v9 :: v_dual_mov_b32 v35, v8
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v36, v7 :: v_dual_mov_b32 v37, v6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, v5 :: v_dual_mov_b32 v39, v4
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, v3 :: v_dual_mov_b32 v49, v2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v1 :: v_dual_mov_b32 v51, v0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s25, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s24, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v11.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v10.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.h, v32.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.h, v32.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v70, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v68, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v65, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v55, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v54, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v52, 0xffff, v11
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s25, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s24, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s23, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s46, s0, 16
+; GFX11-TRUE16-NEXT:    s_mov_b32 s15, 0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s46
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s16, s4
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s17, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s18, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s19, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s20, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s21, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s22, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s23, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s24, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s25, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s26, s14
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s16, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s17, s28, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s18, s29, s40
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB55_4
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_mov_b32 s75, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v37.h
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s75, s75, s9
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s75 :: v_dual_and_b32 v9, 0xffff, v37
-; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s57, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s63, s20
-; GFX11-TRUE16-NEXT:    s_mov_b32 s73, s22
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s47, s47, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s57, s57, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s63, s63, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s73, s73, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v38.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s73 :: v_dual_and_b32 v7, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v32
-; GFX11-TRUE16-NEXT:    s_mov_b32 s59, s16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s61, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s59, s59, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s61, s61, s15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s63
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v34
-; GFX11-TRUE16-NEXT:    s_mov_b32 s77, s26
-; GFX11-TRUE16-NEXT:    s_mov_b32 s79, s28
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s7
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s79, s79, s5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, s59
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v36.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s77 :: v_dual_and_b32 v1, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s61
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v35.h
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s79 :: v_dual_and_b32 v3, 0xffff, v35
-; GFX11-TRUE16-NEXT:    s_mov_b32 s56, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s58, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s60, s17
-; GFX11-TRUE16-NEXT:    s_mov_b32 s62, s19
-; GFX11-TRUE16-NEXT:    s_mov_b32 s72, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s74, s23
-; GFX11-TRUE16-NEXT:    s_mov_b32 s76, s25
-; GFX11-TRUE16-NEXT:    s_mov_b32 s78, s27
-; GFX11-TRUE16-NEXT:    s_mov_b32 s88, s29
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s56, s56, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s58, s58, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s60, s60, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s62, s62, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s72, s72, s12
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s74, s74, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s78, s78, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s88, s88, s4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s47 :: v_dual_mov_b32 v1, s56
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s57 :: v_dual_mov_b32 v3, s58
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, s60
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, s62
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, s72
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, s74
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v13, s76
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v15, s78
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v17, s88
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v7, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s8 :: v_dual_mov_b32 v9, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s10 :: v_dual_mov_b32 v11, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s12 :: v_dual_mov_b32 v13, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s14 :: v_dual_mov_b32 v15, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s17 :: v_dual_mov_b32 v17, s18
+; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s15
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB55_3
 ; GFX11-TRUE16-NEXT:  .LBB55_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s44
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v51, 16, v71
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v50, 16, v70
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v49, 16, v69
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v48, 16, v68
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v39, 16, v67
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v37, 16, v65
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v36, 16, v64
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v35, 16, v55
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v34, 16, v54
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v33, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v52
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s2
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s43
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s3
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s42
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s16
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s41
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s40
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v51.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v51
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s15
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s20
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v50.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v50
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v16, 16, v17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s22
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v20 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v49.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v49
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v48.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v48
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v39.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v38.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v38
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v37.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s24
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s9
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s25
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v22, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v25
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v26, 16, v27
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v28, 16, v29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v36.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v35.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v35
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v34.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v34
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v33.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v32.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s27
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v27, 16, v28
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v29, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v31, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v34, 16, v32
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s1, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s29
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s4 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, s5 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, s7 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s17 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s18 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
@@ -51364,25 +49206,49 @@ define inreg <60 x half> @bitcast_v60i16_to_v60f16_scalar(<60 x i16> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v60i16_to_v60f16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s25, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v11.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v10.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v29.h
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB57_3
@@ -51390,59 +49256,48 @@ define inreg <60 x half> @bitcast_v60i16_to_v60f16_scalar(<60 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB57_4
 ; GFX11-TRUE16-NEXT:  .LBB57_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v9.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v11.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v10.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v14, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v6.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v12, 16, v11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v8.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v13, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v14, 16, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v1.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v7.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v5.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v4.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v14, 16, v1
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s15, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v13, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v15, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v12, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v2.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v29, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v28, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v27, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v26, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v25, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s15, s11
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, s14, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, s11, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v12, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v13, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v15, 16, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s14, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s11, s9
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, s10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, s9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s10, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s9, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s16, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
@@ -51455,194 +49310,146 @@ define inreg <60 x half> @bitcast_v60i16_to_v60f16_scalar(<60 x i16> inreg %a, i
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v22, s8, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v23, s7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s8, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s7, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, s29, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, s28, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, s27, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, s26, 3 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, s25, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v24, s6, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v37, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v36, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v32, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v31, s3, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v30, s5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v64.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v65.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v66.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v67.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v68.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v69.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v70.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v71.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v80.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v81.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v82.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v83.l
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, s15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v34, s14, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v35, s13, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v30, s12, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v31, s11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v32, s10, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v33, s9, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v49, s8, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v48, s0, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v39, s1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v38, s2, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v37, s3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_pk_add_u16 v36, s6, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v31
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v30
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_5
 ; GFX11-TRUE16-NEXT:  .LBB57_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB57_2
 ; GFX11-TRUE16-NEXT:  .LBB57_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s29 :: v_dual_mov_b32 v16, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s27 :: v_dual_mov_b32 v12, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s25 :: v_dual_mov_b32 v14, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s23 :: v_dual_mov_b32 v35, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s21 :: v_dual_mov_b32 v31, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s19 :: v_dual_mov_b32 v33, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, s17 :: v_dual_mov_b32 v36, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, s3 :: v_dual_mov_b32 v38, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, s1 :: v_dual_mov_b32 v48, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, s44 :: v_dual_mov_b32 v51, s43
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s42 :: v_dual_mov_b32 v53, s41
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, s40 :: v_dual_mov_b32 v55, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, s14 :: v_dual_mov_b32 v65, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v66, s12 :: v_dual_mov_b32 v67, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v68, s10 :: v_dual_mov_b32 v69, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v70, s8 :: v_dual_mov_b32 v71, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v80, s5 :: v_dual_mov_b32 v81, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v82, s4 :: v_dual_mov_b32 v83, s45
 ; GFX11-TRUE16-NEXT:  .LBB57_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v38, 16, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v39, 16, v64
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v39, v49, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v35, 16, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v39
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v49, v70, 16, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v84, 0xffff, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v82, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v39, v80, 16, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v3
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v38, v81, 16, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v85, 0xffff, v36
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v29, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v49, v33, 16, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v38, v48, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v54, 16, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v80, 0xffff, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v2
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v83, 16, v48
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v48, v71, 16, v85
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff, v31
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v69, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v68, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v66, 16, v80
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v67, 16, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v34
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v48, v55, 16, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v65, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v64, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v55, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v53, 16, v67
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v50, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v52, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v28, 16, v19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v0.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v1.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v27, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v2.h
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v3.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v7.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v10.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_and_b32 v13, 0xffff, v13
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v9.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v53, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v8.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.h
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v51, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v26, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v36 :: v_dual_mov_b32 v1, v37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v48 :: v_dual_mov_b32 v5, v49
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_mov_b32 v8, v32
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, v33
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v54, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v52, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v51, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v50, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v4
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v36
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v48
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v30 :: v_dual_mov_b32 v7, v31
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v32 :: v_dual_mov_b32 v9, v33
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: bitcast_v60i16_to_v60f16_scalar:
@@ -54259,25 +52066,49 @@ define inreg <60 x i16> @bitcast_v60f16_to_v60i16_scalar(<60 x half> inreg %a, i
 ; GFX11-TRUE16-LABEL: bitcast_v60f16_to_v60i16_scalar:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, 0
 ; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v12
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s29, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s28, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s27, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s25, 16
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v11.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v10.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v9.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v8.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v7.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v3.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v2.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v29.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v0.h
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v29.h
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s44, s29, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s43, s28, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s41, s26, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s25, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s15, s24, 16
 ; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s23, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s22, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s21, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s20, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s19, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s18, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s17, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s16, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s3, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s40, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s0, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s22, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s21, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s20, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s19, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s18, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s17, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s6, s16, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s7, s2, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s4, s1, 16
+; GFX11-TRUE16-NEXT:    s_lshr_b32 s45, s0, 16
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s46, 0
 ; GFX11-TRUE16-NEXT:    s_and_b32 s47, vcc_lo, exec_lo
 ; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB59_3
@@ -54285,59 +52116,48 @@ define inreg <60 x i16> @bitcast_v60f16_to_v60i16_scalar(<60 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s46
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB59_4
 ; GFX11-TRUE16-NEXT:  .LBB59_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v9.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v11.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v10.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v14, 16, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v6.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v12, 16, v11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v8.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v13, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v14, 16, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v1.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v7.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v5.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v4.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v14, 16, v1
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s15 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s22
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v13, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v15, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v12, 16, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v3.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v2.h
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v0.h
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v29, 16, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v28, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v27, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v26, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v25, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s44
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s43
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s42
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s41
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s40
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s15, s24, s15
 ; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s14, s23, s14
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s15, s11
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, s14 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s21
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, s11 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s11, s20
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v12, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v13, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v15, 16, v0
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s14, s10
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s11, s9
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, s10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s10, s19
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, s9 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, s18
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s10, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s7, s9, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s13, s22, s13
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s12, s21, s12
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s11, s20, s11
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s10, s19, s10
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s9, s18, s9
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s8, s17, s8
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s16, s6
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s5
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s7
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s45
+; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s4
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
@@ -54350,194 +52170,146 @@ define inreg <60 x i16> @bitcast_v60f16_to_v60i16_scalar(<60 x half> inreg %a, i
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v22, 0x200, s8 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s17
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v23, 0x200, s7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s16
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s29, s29, s45
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s28, s28, s44
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s27, s27, s43
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s26, s26, s42
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s25, s25, s41
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s6, s8, s6
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s7, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s40
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s0, s13
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s12
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, s29 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, s28 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, s27 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, s26 op_sel_hi:[0,1]
 ; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, s25 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v24, 0x200, s6 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v37, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v36, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v32, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v31, 0x200, s3 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v30, 0x200, s5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v0
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v6
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v30
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v22
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v20
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v12
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v64.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v65.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v66.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v67.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v68.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v69.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v70.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v71.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v80.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v81.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v82.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v83.l
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, s15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v34, 0x200, s14 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v35, 0x200, s13 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v30, 0x200, s12 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v31, 0x200, s11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v32, 0x200, s10 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v33, 0x200, s9 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v49, 0x200, s8 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v48, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v39, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v38, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v37, 0x200, s3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v36, 0x200, s6 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v48
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v38
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v37
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v36
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v49
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v31
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v30
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v34
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v15
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v5
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v6
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v7
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v8
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v9
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v10
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v11
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_5
 ; GFX11-TRUE16-NEXT:  .LBB59_3:
 ; GFX11-TRUE16-NEXT:    s_branch .LBB59_2
 ; GFX11-TRUE16-NEXT:  .LBB59_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, s29
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, s28
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, s27
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s26
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, s25
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, s24
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, s23
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, s22
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, s21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, s20
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, s19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, s18
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, s17
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, s16
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, s45
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, s44
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, s43
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, s42
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, s41
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, s15
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, s7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, s5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, s40
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s29 :: v_dual_mov_b32 v16, s28
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, s27 :: v_dual_mov_b32 v12, s26
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v13, s25 :: v_dual_mov_b32 v14, s24
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v34, s23 :: v_dual_mov_b32 v35, s22
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s21 :: v_dual_mov_b32 v31, s20
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v32, s19 :: v_dual_mov_b32 v33, s18
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, s17 :: v_dual_mov_b32 v36, s16
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, s3 :: v_dual_mov_b32 v38, s2
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, s1 :: v_dual_mov_b32 v48, s0
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, s44 :: v_dual_mov_b32 v51, s43
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, s42 :: v_dual_mov_b32 v53, s41
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, s40 :: v_dual_mov_b32 v55, s15
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, s14 :: v_dual_mov_b32 v65, s13
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v66, s12 :: v_dual_mov_b32 v67, s11
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v68, s10 :: v_dual_mov_b32 v69, s9
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v70, s8 :: v_dual_mov_b32 v71, s6
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v80, s5 :: v_dual_mov_b32 v81, s7
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v82, s4 :: v_dual_mov_b32 v83, s45
 ; GFX11-TRUE16-NEXT:  .LBB59_5: ; %end
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v37, 0xffff, v37
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v64, 0xffff, v36
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v38, 16, v37
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v39, 16, v64
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v39, v49, 16, v31
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v35, 16, v22
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v49, 0xffff, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v39, 0xffff, v39
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v49, v70, 16, v49
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v84, 0xffff, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v1
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v5
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v37, v82, 16, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v18, 16, v0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v4
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v3
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v10
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v5, v49
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v39, v80, 16, v84
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v2
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v9
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v28, 16, v3
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, v39
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v38, v81, 16, v38
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v85, 0xffff, v36
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v29, 16, v18
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v49, v33, 16, v24
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v38, v48, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v54, 16, v20
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v80, 0xffff, v30
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v2
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, v37
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v36, v83, 16, v48
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, v38
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v48, v71, 16, v85
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff, v31
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v69, 16, v33
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v68, 16, v32
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v33, v66, 16, v80
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v67, 16, v71
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v66, 0xffff, v34
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v12
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v48, v55, 16, v30
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v67, 0xffff, v12
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v6
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v65, 16, v35
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v35, v64, 16, v66
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v55, 16, v14
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v53, 16, v67
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v50, 16, v14
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v52, 16, v20
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v34, 16, v23
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v34, v28, 16, v19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v0.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v1.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v25, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v27, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v2.h
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v20, 16, v1
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v6.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v3.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v4.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v5.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v7.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v10.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v10
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_and_b32 v13, 0xffff, v13
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v22, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v5
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v9.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff, v9
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff, v11
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v6, 16, v7
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, v30
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v32, v53, 16, v21
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v8.h
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.h
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v51, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v26, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v0, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v4, 16, v5
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v36 :: v_dual_mov_b32 v1, v37
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v38 :: v_dual_mov_b32 v3, v39
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v48 :: v_dual_mov_b32 v5, v49
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, v31 :: v_dual_mov_b32 v8, v32
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v9, v33
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v11, v35
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v53, 0xffff, v15
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v24, 16, v4
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v7
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v11
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v54, 16, v13
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v52, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v51, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v50, 16, v53
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v4
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v36
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v48
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v30 :: v_dual_mov_b32 v7, v31
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v32 :: v_dual_mov_b32 v9, v33
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v34 :: v_dual_mov_b32 v11, v35
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: bitcast_v60f16_to_v60i16_scalar:
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll
index f888f4f3b1407..5163539046bb0 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll
@@ -789,113 +789,56 @@ define inreg <12 x i8> @bitcast_v3i32_to_v12i8_scalar(<3 x i32> inreg %a, i32 in
 ; GFX9-NEXT:    ; implicit-def: $sgpr6
 ; GFX9-NEXT:    s_branch .LBB5_2
 ;
-; GFX11-TRUE16-LABEL: bitcast_v3i32_to_v12i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB5_4
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB5_3
-; GFX11-TRUE16-NEXT:  .LBB5_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:  .LBB5_3: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s4
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s6
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-TRUE16-NEXT:  .LBB5_4:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr3_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB5_2
-;
-; GFX11-FAKE16-LABEL: bitcast_v3i32_to_v12i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB5_4
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB5_3
-; GFX11-FAKE16-NEXT:  .LBB5_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s1, s1, 3
-; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:  .LBB5_3: ; %end
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s12 :: v_dual_mov_b32 v3, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s11
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s10 :: v_dual_mov_b32 v7, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s3 :: v_dual_mov_b32 v11, s6
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB5_4:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr3
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB5_2
+; GFX11-LABEL: bitcast_v3i32_to_v12i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX11-NEXT:    s_mov_b32 s14, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB5_4
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s3, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s8, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s9, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s10, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s12, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s14
+; GFX11-NEXT:    s_cbranch_vccnz .LBB5_3
+; GFX11-NEXT:  .LBB5_2: ; %cmp.true
+; GFX11-NEXT:    s_add_i32 s2, s2, 3
+; GFX11-NEXT:    s_add_i32 s1, s1, 3
+; GFX11-NEXT:    s_add_i32 s0, s0, 3
+; GFX11-NEXT:    s_lshr_b32 s3, s2, 16
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_lshr_b32 s8, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s9, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s10, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s11, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s12, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:  .LBB5_3: ; %end
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s13
+; GFX11-NEXT:    v_dual_mov_b32 v2, s12 :: v_dual_mov_b32 v3, s4
+; GFX11-NEXT:    v_dual_mov_b32 v4, s1 :: v_dual_mov_b32 v5, s11
+; GFX11-NEXT:    v_dual_mov_b32 v6, s10 :: v_dual_mov_b32 v7, s9
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s8
+; GFX11-NEXT:    v_dual_mov_b32 v10, s3 :: v_dual_mov_b32 v11, s6
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB5_4:
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr3
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB5_2
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -3981,126 +3924,62 @@ define inreg <12 x i8> @bitcast_v3f32_to_v12i8_scalar(<3 x float> inreg %a, i32
 ; GFX9-NEXT:    v_mov_b32_e32 v4, v14
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v3f32_to_v12i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB21_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB21_4
-; GFX11-TRUE16-NEXT:  .LBB21_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, s2, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, s1, 1.0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, s0, 1.0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[12:13], 24, v[8:9]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB21_5
-; GFX11-TRUE16-NEXT:  .LBB21_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB21_2
-; GFX11-TRUE16-NEXT:  .LBB21_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s4
-; GFX11-TRUE16-NEXT:  .LBB21_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v12.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v3f32_to_v12i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s3, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB21_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB21_4
-; GFX11-FAKE16-NEXT:  .LBB21_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, s2, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, s1, 1.0
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, s0, 1.0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[8:9]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v13
-; GFX11-FAKE16-NEXT:    s_branch .LBB21_5
-; GFX11-FAKE16-NEXT:  .LBB21_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB21_2
-; GFX11-FAKE16-NEXT:  .LBB21_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v1, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s12 :: v_dual_mov_b32 v5, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s11 :: v_dual_mov_b32 v7, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s9 :: v_dual_mov_b32 v10, s8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB21_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v14
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v3f32_to_v12i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX11-NEXT:    s_mov_b32 s3, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB21_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s8, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s10, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s11, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s12, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
+; GFX11-NEXT:    s_cbranch_vccnz .LBB21_4
+; GFX11-NEXT:  .LBB21_2: ; %cmp.true
+; GFX11-NEXT:    v_add_f32_e64 v8, s2, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v14, s1, 1.0
+; GFX11-NEXT:    v_add_f32_e64 v13, s0, 1.0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v14
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[8:9]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v13
+; GFX11-NEXT:    s_branch .LBB21_5
+; GFX11-NEXT:  .LBB21_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB21_2
+; GFX11-NEXT:  .LBB21_4:
+; GFX11-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v1, s14
+; GFX11-NEXT:    v_dual_mov_b32 v2, s12 :: v_dual_mov_b32 v5, s13
+; GFX11-NEXT:    v_dual_mov_b32 v6, s11 :: v_dual_mov_b32 v7, s10
+; GFX11-NEXT:    v_dual_mov_b32 v9, s9 :: v_dual_mov_b32 v10, s8
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB21_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v13
+; GFX11-NEXT:    v_mov_b32_e32 v4, v14
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -8396,249 +8275,124 @@ define inreg <12 x i8> @bitcast_v6bf16_to_v12i8_scalar(<6 x bfloat> inreg %a, i3
 ; GFX9-NEXT:    v_mov_b32_e32 v4, s17
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v6bf16_to_v12i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB39_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB39_4
-; GFX11-TRUE16-NEXT:  .LBB39_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s3, 0, s1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s3, 0, s0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s2
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s3
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v7
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s2
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v9
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v14, 0x7fc07fc0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v7, 16, 1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v6, vcc_lo
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v9, 16, 1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v10, v5
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v2, v7
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v8
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v9
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v11 :: v_dual_add_nc_u32 v0, 0x7fff, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v8, 16, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v0, v12, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v2, v10 :: v_dual_add_nc_u32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v7, vcc_lo
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff, v4
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff, v0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v6, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v5, 16, v7
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v9, 16, v10
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[1:2]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v13
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[12:13], 24, v[13:14]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-TRUE16-NEXT:    s_branch .LBB39_5
-; GFX11-TRUE16-NEXT:  .LBB39_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB39_2
-; GFX11-TRUE16-NEXT:  .LBB39_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s4
-; GFX11-TRUE16-NEXT:  .LBB39_5: ; %end
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v12.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v6bf16_to_v12i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s3, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB39_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB39_4
-; GFX11-FAKE16-NEXT:  .LBB39_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s3, 0, s1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s3, 0, s0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s1, 0, s2
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s3
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s2
-; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v4, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, 0x7fc07fc0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v9, vcc_lo
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v0, v8, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v2, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v0, v8
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v10, v4
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v5
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v10, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v0, v11, vcc_lo
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v7, 16, 1
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v9 :: v_dual_add_nc_u32 v3, v3, v7
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v13
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v6, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v4, 16, v5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v7, 16, v9
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[1:2]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[11:12]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v13
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
-; GFX11-FAKE16-NEXT:  .LBB39_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB39_2
-; GFX11-FAKE16-NEXT:  .LBB39_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s12
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s14 :: v_dual_mov_b32 v7, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v5, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s11 :: v_dual_mov_b32 v11, s6
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_mov_b32 v4, s1
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v6bf16_to_v12i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX11-NEXT:    s_mov_b32 s3, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB39_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s13, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s12, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s8, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s14, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s11, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s10, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
+; GFX11-NEXT:    s_cbranch_vccnz .LBB39_4
+; GFX11-NEXT:  .LBB39_2: ; %cmp.true
+; GFX11-NEXT:    s_pack_lh_b32_b16 s3, 0, s1
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    v_add_f32_e64 v0, 0x40c00000, s3
+; GFX11-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
+; GFX11-NEXT:    s_pack_lh_b32_b16 s3, 0, s0
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    s_pack_lh_b32_b16 s1, 0, s2
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_bfe_u32 v3, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v1
+; GFX11-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v1
+; GFX11-NEXT:    s_lshl_b32 s2, s2, 16
+; GFX11-NEXT:    v_add_f32_e64 v4, 0x40c00000, s3
+; GFX11-NEXT:    v_add_f32_e64 v8, 0x40c00000, s2
+; GFX11-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    v_bfe_u32 v10, v4, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v11, 0x400000, v8
+; GFX11-NEXT:    v_mov_b32_e32 v12, 0x7fc07fc0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_bfe_u32 v2, v5, 16, 1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v3, v9, vcc_lo
+; GFX11-NEXT:    v_bfe_u32 v0, v8, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 16, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v2, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, v0, v8
+; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v10, v4
+; GFX11-NEXT:    v_or_b32_e32 v10, 0x400000, v5
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v10, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-NEXT:    v_cndmask_b32_e32 v8, v0, v11, vcc_lo
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
+; GFX11-NEXT:    v_bfe_u32 v3, v7, 16, 1
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v2, v9 :: v_dual_add_nc_u32 v3, v3, v7
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-NEXT:    v_and_b32_e32 v9, 0xffff, v8
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v13
+; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_lshl_or_b32 v2, v6, 16, v3
+; GFX11-NEXT:    v_lshl_or_b32 v1, v4, 16, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_lshl_or_b32 v11, v7, 16, v9
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[1:2]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[11:12]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
+; GFX11-NEXT:    v_mov_b32_e32 v4, v13
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-NEXT:  .LBB39_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB39_2
+; GFX11-NEXT:  .LBB39_4:
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v9, s12
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s10
+; GFX11-NEXT:    v_dual_mov_b32 v6, s14 :: v_dual_mov_b32 v7, s8
+; GFX11-NEXT:    v_dual_mov_b32 v10, s13 :: v_dual_mov_b32 v5, s9
+; GFX11-NEXT:    v_dual_mov_b32 v2, s11 :: v_dual_mov_b32 v11, s6
+; GFX11-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_mov_b32 v4, s1
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -9975,132 +9729,65 @@ define inreg <12 x i8> @bitcast_v6f16_to_v12i8_scalar(<6 x half> inreg %a, i32 i
 ; GFX9-NEXT:    v_mov_b32_e32 v9, v13
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v6f16_to_v12i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB43_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB43_4
-; GFX11-TRUE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v12, 0x7e007e00
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[8:9], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[12:13], 24, v[11:12]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v11
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB43_5
-; GFX11-TRUE16-NEXT:  .LBB43_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB43_2
-; GFX11-TRUE16-NEXT:  .LBB43_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s4
-; GFX11-TRUE16-NEXT:  .LBB43_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v12.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v6f16_to_v12i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s3, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB43_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB43_4
-; GFX11-FAKE16-NEXT:  .LBB43_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v8, 0x200, s2 op_sel_hi:[0,1]
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v9, 0x7e007e00
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[14:15]
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[8:9]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v15
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v14
-; GFX11-FAKE16-NEXT:    s_branch .LBB43_5
-; GFX11-FAKE16-NEXT:  .LBB43_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB43_2
-; GFX11-FAKE16-NEXT:  .LBB43_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v1, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s12 :: v_dual_mov_b32 v5, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s11 :: v_dual_mov_b32 v7, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s9 :: v_dual_mov_b32 v10, s8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB43_5: ; %end
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, v14 :: v_dual_mov_b32 v9, v13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v15
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v6f16_to_v12i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX11-NEXT:    s_mov_b32 s3, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB43_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s8, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s10, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s11, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s12, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
+; GFX11-NEXT:    s_cbranch_vccnz .LBB43_4
+; GFX11-NEXT:  .LBB43_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s1 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s0 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s2 op_sel_hi:[0,1]
+; GFX11-NEXT:    v_mov_b32_e32 v9, 0x7e007e00
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v15
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[14:15]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[8:9]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v15
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v14
+; GFX11-NEXT:    s_branch .LBB43_5
+; GFX11-NEXT:  .LBB43_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB43_2
+; GFX11-NEXT:  .LBB43_4:
+; GFX11-NEXT:    v_dual_mov_b32 v14, s0 :: v_dual_mov_b32 v15, s1
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v1, s14
+; GFX11-NEXT:    v_dual_mov_b32 v2, s12 :: v_dual_mov_b32 v5, s13
+; GFX11-NEXT:    v_dual_mov_b32 v6, s11 :: v_dual_mov_b32 v7, s10
+; GFX11-NEXT:    v_dual_mov_b32 v13, s9 :: v_dual_mov_b32 v10, s8
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB43_5: ; %end
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v0, v14 :: v_dual_mov_b32 v9, v13
+; GFX11-NEXT:    v_mov_b32_e32 v4, v15
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
@@ -11472,126 +11159,62 @@ define inreg <12 x i8> @bitcast_v6i16_to_v12i8_scalar(<6 x i16> inreg %a, i32 in
 ; GFX9-NEXT:    v_mov_b32_e32 v4, v14
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: bitcast_v6i16_to_v12i8_scalar:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB47_3
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s2, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s9, s2, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s10, s1, 24
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s11, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s12, s1, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s13, s0, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s14, s0, 8
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-TRUE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB47_4
-; GFX11-TRUE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, s2, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, s1, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, s0, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[12:13], 24, v[8:9]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    s_branch .LBB47_5
-; GFX11-TRUE16-NEXT:  .LBB47_3:
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr14_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr13_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr4_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr12_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr11_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr10_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr9_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr8_lo16
-; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr6_lo16
-; GFX11-TRUE16-NEXT:    s_branch .LBB47_2
-; GFX11-TRUE16-NEXT:  .LBB47_4:
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, s0
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, s1
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, s2
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, s14
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, s13
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, s12
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, s11
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, s10
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, s9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, s8
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, s6
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, s4
-; GFX11-TRUE16-NEXT:  .LBB47_5: ; %end
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v11.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v12.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: bitcast_v6i16_to_v12i8_scalar:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s3, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB47_3
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s8, s2, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s9, s2, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s10, s1, 24
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s11, s1, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s13, s1, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s12, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s14, s0, 8
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
-; GFX11-FAKE16-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB47_4
-; GFX11-FAKE16-NEXT:  .LBB47_2: ; %cmp.true
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v8, s2, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v14, s1, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[13:14]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[11:12], 24, v[8:9]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v14
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v13
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v13
-; GFX11-FAKE16-NEXT:    s_branch .LBB47_5
-; GFX11-FAKE16-NEXT:  .LBB47_3:
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr14
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr12
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr4
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr13
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr11
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr10
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr9
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr8
-; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr6
-; GFX11-FAKE16-NEXT:    s_branch .LBB47_2
-; GFX11-FAKE16-NEXT:  .LBB47_4:
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v1, s14
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s12 :: v_dual_mov_b32 v5, s13
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s11 :: v_dual_mov_b32 v7, s10
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s9 :: v_dual_mov_b32 v10, s8
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v11, s6
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s4
-; GFX11-FAKE16-NEXT:  .LBB47_5: ; %end
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v13
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v14
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: bitcast_v6i16_to_v12i8_scalar:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX11-NEXT:    s_mov_b32 s3, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB47_3
+; GFX11-NEXT:  ; %bb.1: ; %cmp.false
+; GFX11-NEXT:    s_lshr_b32 s8, s2, 16
+; GFX11-NEXT:    s_lshr_b32 s9, s2, 8
+; GFX11-NEXT:    s_lshr_b32 s10, s1, 24
+; GFX11-NEXT:    s_lshr_b32 s11, s1, 16
+; GFX11-NEXT:    s_lshr_b32 s13, s1, 8
+; GFX11-NEXT:    s_lshr_b32 s12, s0, 16
+; GFX11-NEXT:    s_lshr_b32 s14, s0, 8
+; GFX11-NEXT:    s_lshr_b64 s[6:7], s[2:3], 24
+; GFX11-NEXT:    s_lshr_b64 s[4:5], s[0:1], 24
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s3
+; GFX11-NEXT:    s_cbranch_vccnz .LBB47_4
+; GFX11-NEXT:  .LBB47_2: ; %cmp.true
+; GFX11-NEXT:    v_pk_add_u16 v8, s2, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v14, s1, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    v_pk_add_u16 v13, s0, 3 op_sel_hi:[1,0]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v8
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[13:14]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v14
+; GFX11-NEXT:    v_lshrrev_b64 v[11:12], 24, v[8:9]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v14
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v13
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v13
+; GFX11-NEXT:    s_branch .LBB47_5
+; GFX11-NEXT:  .LBB47_3:
+; GFX11-NEXT:    ; implicit-def: $sgpr14
+; GFX11-NEXT:    ; implicit-def: $sgpr12
+; GFX11-NEXT:    ; implicit-def: $sgpr4
+; GFX11-NEXT:    ; implicit-def: $sgpr13
+; GFX11-NEXT:    ; implicit-def: $sgpr11
+; GFX11-NEXT:    ; implicit-def: $sgpr10
+; GFX11-NEXT:    ; implicit-def: $sgpr9
+; GFX11-NEXT:    ; implicit-def: $sgpr8
+; GFX11-NEXT:    ; implicit-def: $sgpr6
+; GFX11-NEXT:    s_branch .LBB47_2
+; GFX11-NEXT:  .LBB47_4:
+; GFX11-NEXT:    v_dual_mov_b32 v13, s0 :: v_dual_mov_b32 v14, s1
+; GFX11-NEXT:    v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v1, s14
+; GFX11-NEXT:    v_dual_mov_b32 v2, s12 :: v_dual_mov_b32 v5, s13
+; GFX11-NEXT:    v_dual_mov_b32 v6, s11 :: v_dual_mov_b32 v7, s10
+; GFX11-NEXT:    v_dual_mov_b32 v9, s9 :: v_dual_mov_b32 v10, s8
+; GFX11-NEXT:    v_mov_b32_e32 v11, s6
+; GFX11-NEXT:    v_mov_b32_e32 v3, s4
+; GFX11-NEXT:  .LBB47_5: ; %end
+; GFX11-NEXT:    v_mov_b32_e32 v0, v13
+; GFX11-NEXT:    v_mov_b32_e32 v4, v14
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq i32 %b, 0
   br i1 %cmp, label %cmp.true, label %cmp.false
 
diff --git a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_global_pointer.ll b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_global_pointer.ll
index f979d01e495ba..0cd43c99c054c 100644
--- a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_global_pointer.ll
+++ b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_global_pointer.ll
@@ -7312,9 +7312,8 @@ define amdgpu_kernel void @uniform_or_i8(ptr addrspace(1) %result, ptr addrspace
 ; GFX1164-TRUE16-NEXT:  ; %bb.1:
 ; GFX1164-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1164-TRUE16-NEXT:    s_and_b32 s7, s2, 3
-; GFX1164-TRUE16-NEXT:    s_mov_b32 s8, s6
+; GFX1164-TRUE16-NEXT:    s_and_b32 s8, s6, 0xff
 ; GFX1164-TRUE16-NEXT:    s_lshl_b32 s7, s7, 3
-; GFX1164-TRUE16-NEXT:    s_and_b32 s8, s8, 0xff
 ; GFX1164-TRUE16-NEXT:    s_mov_b32 s11, 0x31016000
 ; GFX1164-TRUE16-NEXT:    s_lshl_b32 s9, s8, s7
 ; GFX1164-TRUE16-NEXT:    s_and_b32 s8, s2, -4
@@ -7392,9 +7391,8 @@ define amdgpu_kernel void @uniform_or_i8(ptr addrspace(1) %result, ptr addrspace
 ; GFX1132-TRUE16-NEXT:  ; %bb.1:
 ; GFX1132-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1132-TRUE16-NEXT:    s_and_b32 s6, s2, 3
-; GFX1132-TRUE16-NEXT:    s_mov_b32 s7, s4
+; GFX1132-TRUE16-NEXT:    s_and_b32 s7, s4, 0xff
 ; GFX1132-TRUE16-NEXT:    s_lshl_b32 s6, s6, 3
-; GFX1132-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
 ; GFX1132-TRUE16-NEXT:    s_and_b32 s8, s2, -4
 ; GFX1132-TRUE16-NEXT:    s_lshl_b32 s7, s7, s6
 ; GFX1132-TRUE16-NEXT:    s_mov_b32 s11, 0x31016000
@@ -7472,9 +7470,8 @@ define amdgpu_kernel void @uniform_or_i8(ptr addrspace(1) %result, ptr addrspace
 ; GFX1264-TRUE16-NEXT:  ; %bb.1:
 ; GFX1264-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX1264-TRUE16-NEXT:    s_and_b32 s7, s2, 3
-; GFX1264-TRUE16-NEXT:    s_mov_b32 s8, s6
+; GFX1264-TRUE16-NEXT:    s_and_b32 s8, s6, 0xff
 ; GFX1264-TRUE16-NEXT:    s_lshl_b32 s7, s7, 3
-; GFX1264-TRUE16-NEXT:    s_and_b32 s8, s8, 0xff
 ; GFX1264-TRUE16-NEXT:    s_mov_b32 s11, 0x31016000
 ; GFX1264-TRUE16-NEXT:    s_lshl_b32 s9, s8, s7
 ; GFX1264-TRUE16-NEXT:    s_and_b32 s8, s2, -4
@@ -7552,9 +7549,8 @@ define amdgpu_kernel void @uniform_or_i8(ptr addrspace(1) %result, ptr addrspace
 ; GFX1232-TRUE16-NEXT:  ; %bb.1:
 ; GFX1232-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX1232-TRUE16-NEXT:    s_and_b32 s6, s2, 3
-; GFX1232-TRUE16-NEXT:    s_mov_b32 s7, s4
+; GFX1232-TRUE16-NEXT:    s_and_b32 s7, s4, 0xff
 ; GFX1232-TRUE16-NEXT:    s_lshl_b32 s6, s6, 3
-; GFX1232-TRUE16-NEXT:    s_and_b32 s7, s7, 0xff
 ; GFX1232-TRUE16-NEXT:    s_and_b32 s8, s2, -4
 ; GFX1232-TRUE16-NEXT:    s_lshl_b32 s7, s7, s6
 ; GFX1232-TRUE16-NEXT:    s_mov_b32 s11, 0x31016000
@@ -7932,14 +7928,13 @@ define amdgpu_kernel void @uniform_add_i8(ptr addrspace(1) %result, ptr addrspac
 ; GFX1164-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1164-TRUE16-NEXT:    s_and_b32 s4, s2, -4
 ; GFX1164-TRUE16-NEXT:    s_mov_b32 s5, s3
-; GFX1164-TRUE16-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
-; GFX1164-TRUE16-NEXT:    s_load_b32 s3, s[4:5], 0x0
 ; GFX1164-TRUE16-NEXT:    s_and_b32 s2, s2, 3
-; GFX1164-TRUE16-NEXT:    s_mov_b32 s7, s10
+; GFX1164-TRUE16-NEXT:    s_load_b32 s3, s[4:5], 0x0
+; GFX1164-TRUE16-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
 ; GFX1164-TRUE16-NEXT:    s_lshl_b32 s11, s2, 3
-; GFX1164-TRUE16-NEXT:    s_mul_i32 s7, s7, s6
+; GFX1164-TRUE16-NEXT:    s_mul_i32 s2, s10, s6
 ; GFX1164-TRUE16-NEXT:    s_lshl_b32 s12, 0xff, s11
-; GFX1164-TRUE16-NEXT:    s_and_b32 s2, s7, 0xff
+; GFX1164-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
 ; GFX1164-TRUE16-NEXT:    s_not_b32 s13, s12
 ; GFX1164-TRUE16-NEXT:    s_lshl_b32 s14, s2, s11
 ; GFX1164-TRUE16-NEXT:    s_mov_b32 s7, 0x31016000
@@ -8066,9 +8061,8 @@ define amdgpu_kernel void @uniform_add_i8(ptr addrspace(1) %result, ptr addrspac
 ; GFX1132-TRUE16-NEXT:    s_and_b32 s2, s2, 3
 ; GFX1132-TRUE16-NEXT:    s_load_b32 s7, s[4:5], 0x0
 ; GFX1132-TRUE16-NEXT:    s_bcnt1_i32_b32 s6, s6
-; GFX1132-TRUE16-NEXT:    s_mov_b32 s11, s8
 ; GFX1132-TRUE16-NEXT:    s_lshl_b32 s2, s2, 3
-; GFX1132-TRUE16-NEXT:    s_mul_i32 s6, s11, s6
+; GFX1132-TRUE16-NEXT:    s_mul_i32 s6, s8, s6
 ; GFX1132-TRUE16-NEXT:    s_lshl_b32 s3, 0xff, s2
 ; GFX1132-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
 ; GFX1132-TRUE16-NEXT:    s_not_b32 s11, s3
@@ -8188,16 +8182,14 @@ define amdgpu_kernel void @uniform_add_i8(ptr addrspace(1) %result, ptr addrspac
 ; GFX1264-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX1264-TRUE16-NEXT:    s_and_b32 s4, s2, -4
 ; GFX1264-TRUE16-NEXT:    s_mov_b32 s5, s3
-; GFX1264-TRUE16-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
-; GFX1264-TRUE16-NEXT:    s_load_b32 s3, s[4:5], 0x0
 ; GFX1264-TRUE16-NEXT:    s_and_b32 s2, s2, 3
-; GFX1264-TRUE16-NEXT:    s_mov_b32 s7, s10
+; GFX1264-TRUE16-NEXT:    s_load_b32 s3, s[4:5], 0x0
+; GFX1264-TRUE16-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
 ; GFX1264-TRUE16-NEXT:    s_lshl_b32 s11, s2, 3
 ; GFX1264-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX1264-TRUE16-NEXT:    s_mul_i32 s7, s7, s6
+; GFX1264-TRUE16-NEXT:    s_mul_i32 s2, s10, s6
 ; GFX1264-TRUE16-NEXT:    s_lshl_b32 s12, 0xff, s11
-; GFX1264-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX1264-TRUE16-NEXT:    s_and_b32 s2, s7, 0xff
+; GFX1264-TRUE16-NEXT:    s_and_b32 s2, s2, 0xff
 ; GFX1264-TRUE16-NEXT:    s_not_b32 s13, s12
 ; GFX1264-TRUE16-NEXT:    s_lshl_b32 s14, s2, s11
 ; GFX1264-TRUE16-NEXT:    s_mov_b32 s7, 0x31016000
@@ -8323,10 +8315,9 @@ define amdgpu_kernel void @uniform_add_i8(ptr addrspace(1) %result, ptr addrspac
 ; GFX1232-TRUE16-NEXT:    s_and_b32 s2, s2, 3
 ; GFX1232-TRUE16-NEXT:    s_load_b32 s7, s[4:5], 0x0
 ; GFX1232-TRUE16-NEXT:    s_bcnt1_i32_b32 s6, s6
-; GFX1232-TRUE16-NEXT:    s_mov_b32 s11, s8
 ; GFX1232-TRUE16-NEXT:    s_lshl_b32 s2, s2, 3
 ; GFX1232-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX1232-TRUE16-NEXT:    s_mul_i32 s6, s11, s6
+; GFX1232-TRUE16-NEXT:    s_mul_i32 s6, s8, s6
 ; GFX1232-TRUE16-NEXT:    s_lshl_b32 s3, 0xff, s2
 ; GFX1232-TRUE16-NEXT:    s_wait_alu 0xfffe
 ; GFX1232-TRUE16-NEXT:    s_and_b32 s6, s6, 0xff
@@ -8818,7 +8809,7 @@ define amdgpu_kernel void @uniform_or_i16(ptr addrspace(1) %result, ptr addrspac
 ; GFX7LESS-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
 ; GFX7LESS-NEXT:    v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
 ; GFX7LESS-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
-; GFX7LESS-NEXT:                       ; implicit-def: $vgpr0
+; GFX7LESS-NEXT:    ; implicit-def: $vgpr0
 ; GFX7LESS-NEXT:    s_and_saveexec_b64 s[4:5], vcc
 ; GFX7LESS-NEXT:    s_cbranch_execz .LBB15_2
 ; GFX7LESS-NEXT:  ; %bb.1:
@@ -9010,9 +9001,8 @@ define amdgpu_kernel void @uniform_or_i16(ptr addrspace(1) %result, ptr addrspac
 ; GFX1164-TRUE16-NEXT:  ; %bb.1:
 ; GFX1164-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1164-TRUE16-NEXT:    s_and_b32 s7, s2, 3
-; GFX1164-TRUE16-NEXT:    s_mov_b32 s8, s6
+; GFX1164-TRUE16-NEXT:    s_and_b32 s8, 0xffff, s6
 ; GFX1164-TRUE16-NEXT:    s_lshl_b32 s7, s7, 3
-; GFX1164-TRUE16-NEXT:    s_and_b32 s8, 0xffff, s8
 ; GFX1164-TRUE16-NEXT:    s_mov_b32 s11, 0x31016000
 ; GFX1164-TRUE16-NEXT:    s_lshl_b32 s9, s8, s7
 ; GFX1164-TRUE16-NEXT:    s_and_b32 s8, s2, -4
@@ -9090,9 +9080,8 @@ define amdgpu_kernel void @uniform_or_i16(ptr addrspace(1) %result, ptr addrspac
 ; GFX1132-TRUE16-NEXT:  ; %bb.1:
 ; GFX1132-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1132-TRUE16-NEXT:    s_and_b32 s6, s2, 3
-; GFX1132-TRUE16-NEXT:    s_mov_b32 s7, s4
+; GFX1132-TRUE16-NEXT:    s_and_b32 s7, 0xffff, s4
 ; GFX1132-TRUE16-NEXT:    s_lshl_b32 s6, s6, 3
-; GFX1132-TRUE16-NEXT:    s_and_b32 s7, 0xffff, s7
 ; GFX1132-TRUE16-NEXT:    s_and_b32 s8, s2, -4
 ; GFX1132-TRUE16-NEXT:    s_lshl_b32 s7, s7, s6
 ; GFX1132-TRUE16-NEXT:    s_mov_b32 s11, 0x31016000
@@ -9170,9 +9159,8 @@ define amdgpu_kernel void @uniform_or_i16(ptr addrspace(1) %result, ptr addrspac
 ; GFX1264-TRUE16-NEXT:  ; %bb.1:
 ; GFX1264-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX1264-TRUE16-NEXT:    s_and_b32 s7, s2, 3
-; GFX1264-TRUE16-NEXT:    s_mov_b32 s8, s6
+; GFX1264-TRUE16-NEXT:    s_and_b32 s8, 0xffff, s6
 ; GFX1264-TRUE16-NEXT:    s_lshl_b32 s7, s7, 3
-; GFX1264-TRUE16-NEXT:    s_and_b32 s8, 0xffff, s8
 ; GFX1264-TRUE16-NEXT:    s_mov_b32 s11, 0x31016000
 ; GFX1264-TRUE16-NEXT:    s_lshl_b32 s9, s8, s7
 ; GFX1264-TRUE16-NEXT:    s_and_b32 s8, s2, -4
@@ -9250,9 +9238,8 @@ define amdgpu_kernel void @uniform_or_i16(ptr addrspace(1) %result, ptr addrspac
 ; GFX1232-TRUE16-NEXT:  ; %bb.1:
 ; GFX1232-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX1232-TRUE16-NEXT:    s_and_b32 s6, s2, 3
-; GFX1232-TRUE16-NEXT:    s_mov_b32 s7, s4
+; GFX1232-TRUE16-NEXT:    s_and_b32 s7, 0xffff, s4
 ; GFX1232-TRUE16-NEXT:    s_lshl_b32 s6, s6, 3
-; GFX1232-TRUE16-NEXT:    s_and_b32 s7, 0xffff, s7
 ; GFX1232-TRUE16-NEXT:    s_and_b32 s8, s2, -4
 ; GFX1232-TRUE16-NEXT:    s_lshl_b32 s7, s7, s6
 ; GFX1232-TRUE16-NEXT:    s_mov_b32 s11, 0x31016000
@@ -9328,7 +9315,7 @@ define amdgpu_kernel void @uniform_add_i16(ptr addrspace(1) %result, ptr addrspa
 ; GFX7LESS-NEXT:    v_mbcnt_lo_u32_b32_e64 v0, s6, 0
 ; GFX7LESS-NEXT:    v_mbcnt_hi_u32_b32_e32 v4, s7, v0
 ; GFX7LESS-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v4
-; GFX7LESS-NEXT:                               ; implicit-def: $vgpr0
+; GFX7LESS-NEXT:    ; implicit-def: $vgpr0
 ; GFX7LESS-NEXT:    s_and_saveexec_b64 s[8:9], vcc
 ; GFX7LESS-NEXT:    s_cbranch_execz .LBB16_4
 ; GFX7LESS-NEXT:  ; %bb.1:
@@ -9630,14 +9617,13 @@ define amdgpu_kernel void @uniform_add_i16(ptr addrspace(1) %result, ptr addrspa
 ; GFX1164-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1164-TRUE16-NEXT:    s_and_b32 s4, s2, -4
 ; GFX1164-TRUE16-NEXT:    s_mov_b32 s5, s3
-; GFX1164-TRUE16-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
-; GFX1164-TRUE16-NEXT:    s_load_b32 s3, s[4:5], 0x0
 ; GFX1164-TRUE16-NEXT:    s_and_b32 s2, s2, 3
-; GFX1164-TRUE16-NEXT:    s_mov_b32 s7, s10
+; GFX1164-TRUE16-NEXT:    s_load_b32 s3, s[4:5], 0x0
+; GFX1164-TRUE16-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
 ; GFX1164-TRUE16-NEXT:    s_lshl_b32 s11, s2, 3
-; GFX1164-TRUE16-NEXT:    s_mul_i32 s7, s7, s6
+; GFX1164-TRUE16-NEXT:    s_mul_i32 s2, s10, s6
 ; GFX1164-TRUE16-NEXT:    s_lshl_b32 s12, 0xffff, s11
-; GFX1164-TRUE16-NEXT:    s_and_b32 s2, s7, 0xffff
+; GFX1164-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
 ; GFX1164-TRUE16-NEXT:    s_not_b32 s13, s12
 ; GFX1164-TRUE16-NEXT:    s_lshl_b32 s14, s2, s11
 ; GFX1164-TRUE16-NEXT:    s_mov_b32 s7, 0x31016000
@@ -9764,9 +9750,8 @@ define amdgpu_kernel void @uniform_add_i16(ptr addrspace(1) %result, ptr addrspa
 ; GFX1132-TRUE16-NEXT:    s_and_b32 s2, s2, 3
 ; GFX1132-TRUE16-NEXT:    s_load_b32 s7, s[4:5], 0x0
 ; GFX1132-TRUE16-NEXT:    s_bcnt1_i32_b32 s6, s6
-; GFX1132-TRUE16-NEXT:    s_mov_b32 s11, s8
 ; GFX1132-TRUE16-NEXT:    s_lshl_b32 s2, s2, 3
-; GFX1132-TRUE16-NEXT:    s_mul_i32 s6, s11, s6
+; GFX1132-TRUE16-NEXT:    s_mul_i32 s6, s8, s6
 ; GFX1132-TRUE16-NEXT:    s_lshl_b32 s3, 0xffff, s2
 ; GFX1132-TRUE16-NEXT:    s_and_b32 s6, s6, 0xffff
 ; GFX1132-TRUE16-NEXT:    s_not_b32 s11, s3
@@ -9886,16 +9871,14 @@ define amdgpu_kernel void @uniform_add_i16(ptr addrspace(1) %result, ptr addrspa
 ; GFX1264-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX1264-TRUE16-NEXT:    s_and_b32 s4, s2, -4
 ; GFX1264-TRUE16-NEXT:    s_mov_b32 s5, s3
-; GFX1264-TRUE16-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
-; GFX1264-TRUE16-NEXT:    s_load_b32 s3, s[4:5], 0x0
 ; GFX1264-TRUE16-NEXT:    s_and_b32 s2, s2, 3
-; GFX1264-TRUE16-NEXT:    s_mov_b32 s7, s10
+; GFX1264-TRUE16-NEXT:    s_load_b32 s3, s[4:5], 0x0
+; GFX1264-TRUE16-NEXT:    s_bcnt1_i32_b64 s6, s[6:7]
 ; GFX1264-TRUE16-NEXT:    s_lshl_b32 s11, s2, 3
 ; GFX1264-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX1264-TRUE16-NEXT:    s_mul_i32 s7, s7, s6
+; GFX1264-TRUE16-NEXT:    s_mul_i32 s2, s10, s6
 ; GFX1264-TRUE16-NEXT:    s_lshl_b32 s12, 0xffff, s11
-; GFX1264-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX1264-TRUE16-NEXT:    s_and_b32 s2, s7, 0xffff
+; GFX1264-TRUE16-NEXT:    s_and_b32 s2, s2, 0xffff
 ; GFX1264-TRUE16-NEXT:    s_not_b32 s13, s12
 ; GFX1264-TRUE16-NEXT:    s_lshl_b32 s14, s2, s11
 ; GFX1264-TRUE16-NEXT:    s_mov_b32 s7, 0x31016000
@@ -10021,10 +10004,9 @@ define amdgpu_kernel void @uniform_add_i16(ptr addrspace(1) %result, ptr addrspa
 ; GFX1232-TRUE16-NEXT:    s_and_b32 s2, s2, 3
 ; GFX1232-TRUE16-NEXT:    s_load_b32 s7, s[4:5], 0x0
 ; GFX1232-TRUE16-NEXT:    s_bcnt1_i32_b32 s6, s6
-; GFX1232-TRUE16-NEXT:    s_mov_b32 s11, s8
 ; GFX1232-TRUE16-NEXT:    s_lshl_b32 s2, s2, 3
 ; GFX1232-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX1232-TRUE16-NEXT:    s_mul_i32 s6, s11, s6
+; GFX1232-TRUE16-NEXT:    s_mul_i32 s6, s8, s6
 ; GFX1232-TRUE16-NEXT:    s_lshl_b32 s3, 0xffff, s2
 ; GFX1232-TRUE16-NEXT:    s_wait_alu 0xfffe
 ; GFX1232-TRUE16-NEXT:    s_and_b32 s6, s6, 0xffff
diff --git a/llvm/test/CodeGen/AMDGPU/br_cc.f16.ll b/llvm/test/CodeGen/AMDGPU/br_cc.f16.ll
index 4cb5b7c43a46d..b27ad26cf97b9 100644
--- a/llvm/test/CodeGen/AMDGPU/br_cc.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/br_cc.f16.ll
@@ -77,20 +77,19 @@ define amdgpu_kernel void @br_cc_f16(
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-TRUE16-NEXT:    buffer_load_u16 v1, off, s[8:11], 0 glc dlc
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v0.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v1.l
+; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s6
+; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s7
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v0.l
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v1.l
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cmp_nlt_f16_e32 vcc_lo, v0.h, v1.h
+; GFX11-TRUE16-NEXT:    v_cmp_nlt_f16_e32 vcc_lo, v2.l, v2.h
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB0_2
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %one
-; GFX11-TRUE16-NEXT:    s_branch .LBB0_3
-; GFX11-TRUE16-NEXT:  .LBB0_2: ; %two
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.l
-; GFX11-TRUE16-NEXT:  .LBB0_3: ; %one
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s7
 ; GFX11-TRUE16-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
 ; GFX11-TRUE16-NEXT:    s_endpgm
+; GFX11-TRUE16-NEXT:  .LBB0_2: ; %two
+; GFX11-TRUE16-NEXT:    buffer_store_b16 v1, off, s[0:3], 0
+; GFX11-TRUE16-NEXT:    s_endpgm
 ;
 ; GFX11-FAKE16-LABEL: br_cc_f16:
 ; GFX11-FAKE16:       ; %bb.0: ; %entry
@@ -193,15 +192,13 @@ define amdgpu_kernel void @br_cc_f16_imm_a(
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s3
 ; GFX11-TRUE16-NEXT:    buffer_load_u16 v0, off, s[4:7], 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v0.l
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v0.l
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cmp_nlt_f16_e32 vcc_lo, 0.5, v0.h
+; GFX11-TRUE16-NEXT:    v_cmp_nlt_f16_e32 vcc_lo, 0.5, v1.l
 ; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB1_2
 ; GFX11-TRUE16-NEXT:  ; %bb.1: ; %one
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, 0x3800
-; GFX11-TRUE16-NEXT:    s_branch .LBB1_3
 ; GFX11-TRUE16-NEXT:  .LBB1_2: ; %two
-; GFX11-TRUE16-NEXT:  .LBB1_3: ; %one
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s6
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s7
 ; GFX11-TRUE16-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
@@ -301,15 +298,13 @@ define amdgpu_kernel void @br_cc_f16_imm_b(
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s3
 ; GFX11-TRUE16-NEXT:    buffer_load_u16 v0, off, s[4:7], 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v0.l
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v0.l
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cmp_ngt_f16_e32 vcc_lo, 0.5, v0.h
-; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB2_2
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %one
-; GFX11-TRUE16-NEXT:    s_branch .LBB2_3
-; GFX11-TRUE16-NEXT:  .LBB2_2: ; %two
+; GFX11-TRUE16-NEXT:    v_cmp_ngt_f16_e32 vcc_lo, 0.5, v1.l
+; GFX11-TRUE16-NEXT:    s_cbranch_vccz .LBB2_2
+; GFX11-TRUE16-NEXT:  ; %bb.1: ; %two
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, 0x3800
-; GFX11-TRUE16-NEXT:  .LBB2_3: ; %one
+; GFX11-TRUE16-NEXT:  .LBB2_2: ; %one
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s6
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s7
 ; GFX11-TRUE16-NEXT:    buffer_store_b16 v0, off, s[0:3], 0
diff --git a/llvm/test/CodeGen/AMDGPU/bswap.ll b/llvm/test/CodeGen/AMDGPU/bswap.ll
index 5730d75568fd7..4787f21e28598 100644
--- a/llvm/test/CodeGen/AMDGPU/bswap.ll
+++ b/llvm/test/CodeGen/AMDGPU/bswap.ll
@@ -303,32 +303,18 @@ define amdgpu_kernel void @test_bswap_i64(ptr addrspace(1) %out, ptr addrspace(1
 ; VI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
-; GFX11-REAL16-LABEL: test_bswap_i64:
-; GFX11-REAL16:       ; %bb.0:
-; GFX11-REAL16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-REAL16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-REAL16-NEXT:    s_load_b64 s[2:3], s[2:3], 0x0
-; GFX11-REAL16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-REAL16-NEXT:    s_mov_b32 s3, 0x31016000
-; GFX11-REAL16-NEXT:    v_perm_b32 v0, 0, s2, 0x10203
-; GFX11-REAL16-NEXT:    s_mov_b32 s2, -1
-; GFX11-REAL16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-REAL16-NEXT:    v_mov_b32_e32 v1, v0
-; GFX11-REAL16-NEXT:    buffer_store_b64 v[0:1], off, s[0:3], 0
-; GFX11-REAL16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: test_bswap_i64:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_load_b64 s[4:5], s[2:3], 0x0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s3, 0x31016000
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, -1
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_perm_b32 v1, 0, s4, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v0, 0, s5, 0x10203
-; GFX11-FAKE16-NEXT:    buffer_store_b64 v[0:1], off, s[0:3], 0
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: test_bswap_i64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_load_b64 s[4:5], s[2:3], 0x0
+; GFX11-NEXT:    s_mov_b32 s3, 0x31016000
+; GFX11-NEXT:    s_mov_b32 s2, -1
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    v_perm_b32 v1, 0, s4, 0x10203
+; GFX11-NEXT:    v_perm_b32 v0, 0, s5, 0x10203
+; GFX11-NEXT:    buffer_store_b64 v[0:1], off, s[0:3], 0
+; GFX11-NEXT:    s_endpgm
   %val = load i64, ptr addrspace(1) %in, align 8
   %bswap = call i64 @llvm.bswap.i64(i64 %val) nounwind readnone
   store i64 %bswap, ptr addrspace(1) %out, align 8
@@ -378,36 +364,20 @@ define amdgpu_kernel void @test_bswap_v2i64(ptr addrspace(1) %out, ptr addrspace
 ; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
-; GFX11-REAL16-LABEL: test_bswap_v2i64:
-; GFX11-REAL16:       ; %bb.0:
-; GFX11-REAL16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-REAL16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-REAL16-NEXT:    s_load_b128 s[4:7], s[2:3], 0x0
-; GFX11-REAL16-NEXT:    s_mov_b32 s3, 0x31016000
-; GFX11-REAL16-NEXT:    s_mov_b32 s2, -1
-; GFX11-REAL16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-REAL16-NEXT:    v_perm_b32 v0, 0, s4, 0x10203
-; GFX11-REAL16-NEXT:    v_perm_b32 v2, 0, s6, 0x10203
-; GFX11-REAL16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-REAL16-NEXT:    v_mov_b32_e32 v1, v0
-; GFX11-REAL16-NEXT:    v_mov_b32_e32 v3, v2
-; GFX11-REAL16-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
-; GFX11-REAL16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: test_bswap_v2i64:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_load_b128 s[4:7], s[2:3], 0x0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s3, 0x31016000
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, -1
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_perm_b32 v3, 0, s6, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v2, 0, s7, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v1, 0, s4, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v0, 0, s5, 0x10203
-; GFX11-FAKE16-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: test_bswap_v2i64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_load_b128 s[4:7], s[2:3], 0x0
+; GFX11-NEXT:    s_mov_b32 s3, 0x31016000
+; GFX11-NEXT:    s_mov_b32 s2, -1
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    v_perm_b32 v3, 0, s6, 0x10203
+; GFX11-NEXT:    v_perm_b32 v2, 0, s7, 0x10203
+; GFX11-NEXT:    v_perm_b32 v1, 0, s4, 0x10203
+; GFX11-NEXT:    v_perm_b32 v0, 0, s5, 0x10203
+; GFX11-NEXT:    buffer_store_b128 v[0:3], off, s[0:3], 0
+; GFX11-NEXT:    s_endpgm
   %val = load <2 x i64>, ptr addrspace(1) %in, align 16
   %bswap = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> %val) nounwind readnone
   store <2 x i64> %bswap, ptr addrspace(1) %out, align 16
@@ -475,49 +445,26 @@ define amdgpu_kernel void @test_bswap_v4i64(ptr addrspace(1) %out, ptr addrspace
 ; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[12:15], 0
 ; VI-NEXT:    s_endpgm
 ;
-; GFX11-REAL16-LABEL: test_bswap_v4i64:
-; GFX11-REAL16:       ; %bb.0:
-; GFX11-REAL16-NEXT:    s_load_b128 s[8:11], s[4:5], 0x24
-; GFX11-REAL16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-REAL16-NEXT:    s_load_b256 s[0:7], s[10:11], 0x0
-; GFX11-REAL16-NEXT:    s_mov_b32 s11, 0x31016000
-; GFX11-REAL16-NEXT:    s_mov_b32 s10, -1
-; GFX11-REAL16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-REAL16-NEXT:    v_perm_b32 v0, 0, s4, 0x10203
-; GFX11-REAL16-NEXT:    v_perm_b32 v2, 0, s6, 0x10203
-; GFX11-REAL16-NEXT:    v_perm_b32 v4, 0, s0, 0x10203
-; GFX11-REAL16-NEXT:    v_perm_b32 v6, 0, s2, 0x10203
-; GFX11-REAL16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-REAL16-NEXT:    v_mov_b32_e32 v1, v0
-; GFX11-REAL16-NEXT:    v_mov_b32_e32 v3, v2
-; GFX11-REAL16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-REAL16-NEXT:    v_mov_b32_e32 v5, v4
-; GFX11-REAL16-NEXT:    v_mov_b32_e32 v7, v6
-; GFX11-REAL16-NEXT:    s_clause 0x1
-; GFX11-REAL16-NEXT:    buffer_store_b128 v[0:3], off, s[8:11], 0 offset:16
-; GFX11-REAL16-NEXT:    buffer_store_b128 v[4:7], off, s[8:11], 0
-; GFX11-REAL16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: test_bswap_v4i64:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b128 s[8:11], s[4:5], 0x24
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_load_b256 s[0:7], s[10:11], 0x0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s11, 0x31016000
-; GFX11-FAKE16-NEXT:    s_mov_b32 s10, -1
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_perm_b32 v7, 0, s6, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v6, 0, s7, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v5, 0, s4, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v4, 0, s5, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v3, 0, s2, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v2, 0, s3, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v1, 0, s0, 0x10203
-; GFX11-FAKE16-NEXT:    v_perm_b32 v0, 0, s1, 0x10203
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    buffer_store_b128 v[4:7], off, s[8:11], 0 offset:16
-; GFX11-FAKE16-NEXT:    buffer_store_b128 v[0:3], off, s[8:11], 0
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: test_bswap_v4i64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b128 s[8:11], s[4:5], 0x24
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_load_b256 s[0:7], s[10:11], 0x0
+; GFX11-NEXT:    s_mov_b32 s11, 0x31016000
+; GFX11-NEXT:    s_mov_b32 s10, -1
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    v_perm_b32 v7, 0, s6, 0x10203
+; GFX11-NEXT:    v_perm_b32 v6, 0, s7, 0x10203
+; GFX11-NEXT:    v_perm_b32 v5, 0, s4, 0x10203
+; GFX11-NEXT:    v_perm_b32 v4, 0, s5, 0x10203
+; GFX11-NEXT:    v_perm_b32 v3, 0, s2, 0x10203
+; GFX11-NEXT:    v_perm_b32 v2, 0, s3, 0x10203
+; GFX11-NEXT:    v_perm_b32 v1, 0, s0, 0x10203
+; GFX11-NEXT:    v_perm_b32 v0, 0, s1, 0x10203
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    buffer_store_b128 v[4:7], off, s[8:11], 0 offset:16
+; GFX11-NEXT:    buffer_store_b128 v[0:3], off, s[8:11], 0
+; GFX11-NEXT:    s_endpgm
   %val = load <4 x i64>, ptr addrspace(1) %in, align 32
   %bswap = call <4 x i64> @llvm.bswap.v4i64(<4 x i64> %val) nounwind readnone
   store <4 x i64> %bswap, ptr addrspace(1) %out, align 32
diff --git a/llvm/test/CodeGen/AMDGPU/call-argument-types.ll b/llvm/test/CodeGen/AMDGPU/call-argument-types.ll
index acf2f8add7670..9f47735a2a7b8 100644
--- a/llvm/test/CodeGen/AMDGPU/call-argument-types.ll
+++ b/llvm/test/CodeGen/AMDGPU/call-argument-types.ll
@@ -5281,69 +5281,37 @@ define amdgpu_kernel void @test_call_external_void_func_v16i8() #0 {
 ; GFX9-NEXT:    s_swappc_b64 s[30:31], s[4:5]
 ; GFX9-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: test_call_external_void_func_v16i8:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_load_b64 s[4:5], s[0:1], 0x0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s7, 0x31016000
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, -1
-; GFX11-TRUE16-NEXT:    s_getpc_b64 s[2:3]
-; GFX11-TRUE16-NEXT:    s_add_u32 s2, s2, external_void_func_v16i8 at rel32@lo+4
-; GFX11-TRUE16-NEXT:    s_addc_u32 s3, s3, external_void_func_v16i8 at rel32@hi+12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s32, 0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    buffer_load_b128 v[16:19], off, s[4:7], 0
-; GFX11-TRUE16-NEXT:    s_mov_b64 s[6:7], s[0:1]
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 24, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v16.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v17.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v18.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v19.l
-; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: test_call_external_void_func_v16i8:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b64 s[4:5], s[0:1], 0x0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s7, 0x31016000
-; GFX11-FAKE16-NEXT:    s_mov_b32 s6, -1
-; GFX11-FAKE16-NEXT:    s_getpc_b64 s[2:3]
-; GFX11-FAKE16-NEXT:    s_add_u32 s2, s2, external_void_func_v16i8 at rel32@lo+4
-; GFX11-FAKE16-NEXT:    s_addc_u32 s3, s3, external_void_func_v16i8 at rel32@hi+12
-; GFX11-FAKE16-NEXT:    s_mov_b32 s32, 0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    buffer_load_b128 v[0:3], off, s[4:7], 0
-; GFX11-FAKE16-NEXT:    s_mov_b64 s[6:7], s[0:1]
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 24, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v16
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, v3 :: v_dual_mov_b32 v3, v18
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, v17
-; GFX11-FAKE16-NEXT:    s_swappc_b64 s[30:31], s[2:3]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: test_call_external_void_func_v16i8:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b64 s[4:5], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s7, 0x31016000
+; GFX11-NEXT:    s_mov_b32 s6, -1
+; GFX11-NEXT:    s_getpc_b64 s[2:3]
+; GFX11-NEXT:    s_add_u32 s2, s2, external_void_func_v16i8 at rel32@lo+4
+; GFX11-NEXT:    s_addc_u32 s3, s3, external_void_func_v16i8 at rel32@hi+12
+; GFX11-NEXT:    s_mov_b32 s32, 0
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    buffer_load_b128 v[0:3], off, s[4:7], 0
+; GFX11-NEXT:    s_mov_b64 s[6:7], s[0:1]
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 8, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 24, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v3
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v16
+; GFX11-NEXT:    v_mov_b32_e32 v8, v2
+; GFX11-NEXT:    v_dual_mov_b32 v12, v3 :: v_dual_mov_b32 v3, v18
+; GFX11-NEXT:    v_mov_b32_e32 v2, v17
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[2:3]
+; GFX11-NEXT:    s_endpgm
 ;
 ; HSA-LABEL: test_call_external_void_func_v16i8:
 ; HSA:       ; %bb.0:
diff --git a/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll b/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll
index 71b1a16c79e69..6cd439999a554 100644
--- a/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll
@@ -279,42 +279,23 @@ define amdgpu_kernel void @s_fabs_v4bf16(ptr addrspace(1) %out, <4 x bfloat> %in
 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: s_fabs_v4bf16:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s3, 0x7fff
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s3, 16
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s2
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s5, s5, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s3, 0x7fff
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s5, s2
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s4, s3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v1, s3
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-TRUE16-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: s_fabs_v4bf16:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s3, 0x7fff
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s3, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s2, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s2, s2, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s5, s5, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s3, 0x7fff
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s5
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s4, s3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v1, s3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-FAKE16-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_fabs_v4bf16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_and_b32 s4, s3, 0x7fff
+; GFX11-NEXT:    s_lshr_b32 s3, s3, 16
+; GFX11-NEXT:    s_lshr_b32 s5, s2, 16
+; GFX11-NEXT:    s_and_b32 s2, s2, 0x7fff
+; GFX11-NEXT:    s_and_b32 s5, s5, 0x7fff
+; GFX11-NEXT:    s_and_b32 s3, s3, 0x7fff
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s5
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s4, s3
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v1, s3
+; GFX11-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
+; GFX11-NEXT:    s_endpgm
   %fabs = call <4 x bfloat> @llvm.fabs.v4bf16(<4 x bfloat> %in)
   store <4 x bfloat> %fabs, ptr addrspace(1) %out
   ret void
@@ -385,50 +366,27 @@ define amdgpu_kernel void @fabs_fold_bf16(ptr addrspace(1) %out, bfloat %in0, bf
 ; GFX9-NEXT:    global_store_short_d16_hi v0, v1, s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: fabs_fold_bf16:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_load_b32 s0, s[4:5], 0x8
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_mul_f32_e64 v0, s1, s0
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_cndmask_b32 v0, v1, v2
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_store_d16_hi_b16 v3, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: fabs_fold_bf16:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b32 s0, s[4:5], 0x8
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s0, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mul_f32_e64 v0, s1, s0
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_cndmask_b32 v0, v1, v2
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_store_d16_hi_b16 v3, v0, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: fabs_fold_bf16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_and_b32 s1, s0, 0x7fff
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mul_f32_e64 v0, s1, s0
+; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
+; GFX11-NEXT:    v_bfe_u32 v1, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_cndmask_b32 v0, v1, v2
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    global_store_d16_hi_b16 v3, v0, s[0:1]
+; GFX11-NEXT:    s_endpgm
   %fabs = call bfloat @llvm.fabs.bf16(bfloat %in0)
   %fmul = fmul bfloat %fabs, %in1
   store bfloat %fmul, ptr addrspace(1) %out
diff --git a/llvm/test/CodeGen/AMDGPU/fcopysign.bf16.ll b/llvm/test/CodeGen/AMDGPU/fcopysign.bf16.ll
index 7c89a41d62fbf..8c59f4434cc93 100644
--- a/llvm/test/CodeGen/AMDGPU/fcopysign.bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fcopysign.bf16.ll
@@ -3682,40 +3682,23 @@ define amdgpu_ps i16 @s_copysign_out_bf16_mag_f32_sign_bf16(float inreg %mag, bf
 ; GFX10-NEXT:    v_readfirstlane_b32 s0, v0
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
-; GFX11TRUE16-LABEL: s_copysign_out_bf16_mag_f32_sign_bf16:
-; GFX11TRUE16:       ; %bb.0:
-; GFX11TRUE16-NEXT:    s_bfe_u32 s2, s0, 0x10010
-; GFX11TRUE16-NEXT:    v_cmp_u_f32_e64 s3, s0, s0
-; GFX11TRUE16-NEXT:    s_add_i32 s2, s2, s0
-; GFX11TRUE16-NEXT:    s_bitset1_b32 s0, 22
-; GFX11TRUE16-NEXT:    s_addk_i32 s2, 0x7fff
-; GFX11TRUE16-NEXT:    s_and_b32 s3, s3, exec_lo
-; GFX11TRUE16-NEXT:    s_cselect_b32 s0, s0, s2
-; GFX11TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11TRUE16-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11TRUE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11TRUE16-NEXT:    v_bfi_b32 v0, 0x7fff, v0, s1
-; GFX11TRUE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11FAKE16-LABEL: s_copysign_out_bf16_mag_f32_sign_bf16:
-; GFX11FAKE16:       ; %bb.0:
-; GFX11FAKE16-NEXT:    s_bfe_u32 s2, s0, 0x10010
-; GFX11FAKE16-NEXT:    v_cmp_u_f32_e64 s3, s0, s0
-; GFX11FAKE16-NEXT:    s_add_i32 s2, s2, s0
-; GFX11FAKE16-NEXT:    s_bitset1_b32 s0, 22
-; GFX11FAKE16-NEXT:    s_addk_i32 s2, 0x7fff
-; GFX11FAKE16-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11FAKE16-NEXT:    s_and_b32 s3, s3, exec_lo
-; GFX11FAKE16-NEXT:    s_cselect_b32 s0, s0, s2
-; GFX11FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11FAKE16-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
-; GFX11FAKE16-NEXT:    v_bfi_b32 v0, 0x7fff, s0, v0
-; GFX11FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11FAKE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11FAKE16-NEXT:    ; return to shader part epilog
+; GFX11-LABEL: s_copysign_out_bf16_mag_f32_sign_bf16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_bfe_u32 s2, s0, 0x10010
+; GFX11-NEXT:    v_cmp_u_f32_e64 s3, s0, s0
+; GFX11-NEXT:    s_add_i32 s2, s2, s0
+; GFX11-NEXT:    s_bitset1_b32 s0, 22
+; GFX11-NEXT:    s_addk_i32 s2, 0x7fff
+; GFX11-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-NEXT:    s_and_b32 s3, s3, exec_lo
+; GFX11-NEXT:    s_cselect_b32 s0, s0, s2
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_lshr_b32 s0, s0, 16
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, s0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    ; return to shader part epilog
   %mag.trunc = fptrunc float %mag to bfloat
   %out = call bfloat @llvm.copysign.bf16(bfloat %mag.trunc, bfloat %sign)
   %cast = bitcast bfloat %out to i16
@@ -5436,27 +5419,15 @@ define amdgpu_ps i32 @s_copysign_out_v2bf16_mag_v2bf16_sign_v2f64(<2 x bfloat> i
 ; GFX10-NEXT:    v_readfirstlane_b32 s0, v0
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
-; GFX11TRUE16-LABEL: s_copysign_out_v2bf16_mag_v2bf16_sign_v2f64:
-; GFX11TRUE16:       ; %bb.0:
-; GFX11TRUE16-NEXT:    s_mov_b32 s1, s2
-; GFX11TRUE16-NEXT:    s_mov_b32 s2, s4
-; GFX11TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s2
-; GFX11TRUE16-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11TRUE16-NEXT:    v_bfi_b32 v0, 0x7fff7fff, s0, v0
-; GFX11TRUE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11FAKE16-LABEL: s_copysign_out_v2bf16_mag_v2bf16_sign_v2f64:
-; GFX11FAKE16:       ; %bb.0:
-; GFX11FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s2, s4
-; GFX11FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11FAKE16-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11FAKE16-NEXT:    v_bfi_b32 v0, 0x7fff7fff, s0, v0
-; GFX11FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11FAKE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11FAKE16-NEXT:    ; return to shader part epilog
+; GFX11-LABEL: s_copysign_out_v2bf16_mag_v2bf16_sign_v2f64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s2, s4
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff7fff, s0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    ; return to shader part epilog
   %sign.trunc = fptrunc <2 x double> %sign to <2 x bfloat>
   %out = call <2 x bfloat> @llvm.copysign.v2bf16(<2 x bfloat> %mag, <2 x bfloat> %sign.trunc)
   %cast = bitcast <2 x bfloat> %out to i32
diff --git a/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll b/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll
index e74d5ba24079d..b7c8f9418b14d 100644
--- a/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll
@@ -275,23 +275,14 @@ define i16 @s_test_copysign_f16_10_mag(half inreg %sign) {
 ; GFX9-NEXT:    v_mov_b32_e32 v0, s4
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: s_test_copysign_f16_10_mag:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0x8000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s0, 0x4900
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s0
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: s_test_copysign_f16_10_mag:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0x8000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, 0x4900
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: s_test_copysign_f16_10_mag:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_and_b32 s0, s0, 0x8000
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s0, s0, 0x4900
+; GFX11-NEXT:    v_mov_b32_e32 v0, s0
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %result = call half @llvm.copysign.f16(half 10.0, half %sign)
   %cast = bitcast half %result to i16
   ret i16 %cast
@@ -1197,114 +1188,59 @@ define amdgpu_ps i16 @s_copysign_out_f16_mag_f64_sign_f16(double inreg %mag, hal
 ; GFX9-NEXT:    v_readfirstlane_b32 s0, v0
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX11-TRUE16-LABEL: s_copysign_out_f16_mag_f64_sign_f16:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_and_b32 s3, s1, 0x1ff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_or_b32 s0, s3, s0
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s0, 0
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s0, -1, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
-; GFX11-TRUE16-NEXT:    s_bfe_u32 s0, s1, 0xb0014
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s1, s1, 8
-; GFX11-TRUE16-NEXT:    s_sub_i32 s3, 0x3f1, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0xffe
-; GFX11-TRUE16-NEXT:    v_med3_i32 v1, s3, 0, 13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s3, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-TRUE16-NEXT:    s_or_b32 s1, s1, s3
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s1, 0x1000
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s5, s3, s4
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s5, s4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s4, s3
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s3, 1, 0
-; GFX11-TRUE16-NEXT:    s_addk_i32 s0, 0xfc10
-; GFX11-TRUE16-NEXT:    s_or_b32 s3, s5, s3
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s0, 12
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s1, s4
-; GFX11-TRUE16-NEXT:    s_cmp_lt_i32 s0, 1
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s3, 7
-; GFX11-TRUE16-NEXT:    s_cmp_gt_i32 s4, 5
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s5, 1, 0
-; GFX11-TRUE16-NEXT:    s_cmp_eq_u32 s4, 3
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s4, 1, 0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s3, 2
-; GFX11-TRUE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_add_i32 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_cmp_lt_i32 s0, 31
-; GFX11-TRUE16-NEXT:    s_movk_i32 s4, 0x7e00
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s3, s3, 0x7c00
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s1, s4, 0x7c00
-; GFX11-TRUE16-NEXT:    s_cmpk_eq_i32 s0, 0x40f
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s0, s1, s3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, 0x7fff, v0, s2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-FAKE16-LABEL: s_copysign_out_f16_mag_f64_sign_f16:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_and_b32 s3, s1, 0x1ff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s3, s0
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s0, 0
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s0, -1, 0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
-; GFX11-FAKE16-NEXT:    s_bfe_u32 s0, s1, 0xb0014
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s1, s1, 8
-; GFX11-FAKE16-NEXT:    s_sub_i32 s3, 0x3f1, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0xffe
-; GFX11-FAKE16-NEXT:    v_med3_i32 v1, s3, 0, 13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s3, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s1, s3
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s1, 0x1000
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s5, s3, s4
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s5, s4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s4, s3
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s3, 1, 0
-; GFX11-FAKE16-NEXT:    s_addk_i32 s0, 0xfc10
-; GFX11-FAKE16-NEXT:    s_or_b32 s3, s5, s3
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s0, 12
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s1, s4
-; GFX11-FAKE16-NEXT:    s_cmp_lt_i32 s0, 1
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s3, 7
-; GFX11-FAKE16-NEXT:    s_cmp_gt_i32 s4, 5
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s5, 1, 0
-; GFX11-FAKE16-NEXT:    s_cmp_eq_u32 s4, 3
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s4, 1, 0
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s3, 2
-; GFX11-FAKE16-NEXT:    s_or_b32 s4, s4, s5
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_add_i32 s3, s3, s4
-; GFX11-FAKE16-NEXT:    s_cmp_lt_i32 s0, 31
-; GFX11-FAKE16-NEXT:    s_movk_i32 s4, 0x7e00
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s3, s3, 0x7c00
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s1, s4, 0x7c00
-; GFX11-FAKE16-NEXT:    s_cmpk_eq_i32 s0, 0x40f
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s0, s1, s3
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fff, s0, v0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11-FAKE16-NEXT:    ; return to shader part epilog
+; GFX11-LABEL: s_copysign_out_f16_mag_f64_sign_f16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_and_b32 s3, s1, 0x1ff
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s0, s3, s0
+; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX11-NEXT:    s_cselect_b32 s0, -1, 0
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX11-NEXT:    s_bfe_u32 s0, s1, 0xb0014
+; GFX11-NEXT:    s_lshr_b32 s1, s1, 8
+; GFX11-NEXT:    s_sub_i32 s3, 0x3f1, s0
+; GFX11-NEXT:    s_and_b32 s1, s1, 0xffe
+; GFX11-NEXT:    v_med3_i32 v1, s3, 0, 13
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v1
+; GFX11-NEXT:    s_or_b32 s1, s1, s3
+; GFX11-NEXT:    s_or_b32 s3, s1, 0x1000
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_lshr_b32 s5, s3, s4
+; GFX11-NEXT:    s_lshl_b32 s4, s5, s4
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_cmp_lg_u32 s4, s3
+; GFX11-NEXT:    s_cselect_b32 s3, 1, 0
+; GFX11-NEXT:    s_addk_i32 s0, 0xfc10
+; GFX11-NEXT:    s_or_b32 s3, s5, s3
+; GFX11-NEXT:    s_lshl_b32 s4, s0, 12
+; GFX11-NEXT:    s_or_b32 s4, s1, s4
+; GFX11-NEXT:    s_cmp_lt_i32 s0, 1
+; GFX11-NEXT:    s_cselect_b32 s3, s3, s4
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_b32 s4, s3, 7
+; GFX11-NEXT:    s_cmp_gt_i32 s4, 5
+; GFX11-NEXT:    s_cselect_b32 s5, 1, 0
+; GFX11-NEXT:    s_cmp_eq_u32 s4, 3
+; GFX11-NEXT:    s_cselect_b32 s4, 1, 0
+; GFX11-NEXT:    s_lshr_b32 s3, s3, 2
+; GFX11-NEXT:    s_or_b32 s4, s4, s5
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_add_i32 s3, s3, s4
+; GFX11-NEXT:    s_cmp_lt_i32 s0, 31
+; GFX11-NEXT:    s_movk_i32 s4, 0x7e00
+; GFX11-NEXT:    s_cselect_b32 s3, s3, 0x7c00
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_cselect_b32 s1, s4, 0x7c00
+; GFX11-NEXT:    s_cmpk_eq_i32 s0, 0x40f
+; GFX11-NEXT:    s_cselect_b32 s0, s1, s3
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff, s0, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    ; return to shader part epilog
   %mag.trunc = fptrunc double %mag to half
   %result = call half @llvm.copysign.f16(half %mag.trunc, half %sign)
   %cast = bitcast half %result to i16
@@ -4411,27 +4347,15 @@ define amdgpu_ps i32 @s_copysign_out_v2f16_mag_v2f16_sign_v2f64(<2 x half> inreg
 ; GFX9-NEXT:    v_readfirstlane_b32 s0, v0
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX11-TRUE16-LABEL: s_copysign_out_v2f16_mag_v2f16_sign_v2f64:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s2
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, 0x7fff7fff, s0, v0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-FAKE16-LABEL: s_copysign_out_v2f16_mag_v2f16_sign_v2f64:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s2, s4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, 0x7fff7fff, s0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11-FAKE16-NEXT:    ; return to shader part epilog
+; GFX11-LABEL: s_copysign_out_v2f16_mag_v2f16_sign_v2f64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s2, s4
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mov_b32_e32 v0, s1
+; GFX11-NEXT:    v_bfi_b32 v0, 0x7fff7fff, s0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    ; return to shader part epilog
   %sign.trunc = fptrunc <2 x double> %sign to <2 x half>
   %out = call <2 x half> @llvm.copysign.v2f16(<2 x half> %mag, <2 x half> %sign.trunc)
   %cast = bitcast <2 x half> %out to i32
diff --git a/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll b/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll
index d189b6d4c1e83..64a9727330cfd 100644
--- a/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll
@@ -70,50 +70,27 @@ define amdgpu_kernel void @fneg_fabs_fadd_bf16(ptr addrspace(1) %out, bfloat %x,
 ; GFX9-NEXT:    global_store_short_d16_hi v0, v1, s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: fneg_fabs_fadd_bf16:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_load_b32 s0, s[4:5], 0x8
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0x7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_sub_f32_e64 v0, s0, s1
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_cndmask_b32 v0, v1, v2
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_store_d16_hi_b16 v3, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: fneg_fabs_fadd_bf16:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b32 s0, s[4:5], 0x8
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s0, 0x7fff
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_sub_f32_e64 v0, s0, s1
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_cndmask_b32 v0, v1, v2
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_store_d16_hi_b16 v3, v0, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: fneg_fabs_fadd_bf16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_and_b32 s1, s0, 0x7fff
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_sub_f32_e64 v0, s0, s1
+; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
+; GFX11-NEXT:    v_bfe_u32 v1, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_cndmask_b32 v0, v1, v2
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    global_store_d16_hi_b16 v3, v0, s[0:1]
+; GFX11-NEXT:    s_endpgm
   %fabs = call bfloat @llvm.fabs.bf16(bfloat %x)
   %fsub = fsub bfloat -0.0, %fabs
   %fadd = fadd bfloat %y, %fsub
@@ -187,50 +164,27 @@ define amdgpu_kernel void @fneg_fabs_fmul_bf16(ptr addrspace(1) %out, bfloat %x,
 ; GFX9-NEXT:    global_store_short_d16_hi v0, v1, s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: fneg_fabs_fmul_bf16:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_load_b32 s0, s[4:5], 0x8
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s0
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GFX11-TRUE16-NEXT:    s_bitset1_b32 s1, 15
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    v_mul_f32_e64 v0, s0, s1
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_cndmask_b32 v0, v1, v2
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_store_d16_hi_b16 v3, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: fneg_fabs_fmul_bf16:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b32 s0, s[4:5], 0x8
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_or_b32 s1, s0, 0x8000
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_mul_f32_e64 v0, s0, s1
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v0
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_cndmask_b32 v0, v1, v2
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_store_d16_hi_b16 v3, v0, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: fneg_fabs_fmul_bf16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_or_b32 s1, s0, 0x8000
+; GFX11-NEXT:    s_and_b32 s0, s0, 0xffff0000
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_mul_f32_e64 v0, s0, s1
+; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
+; GFX11-NEXT:    v_bfe_u32 v1, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v2, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, v1, v0
+; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_cndmask_b32 v0, v1, v2
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    global_store_d16_hi_b16 v3, v0, s[0:1]
+; GFX11-NEXT:    s_endpgm
   %fabs = call bfloat @llvm.fabs.bf16(bfloat %x)
   %fsub = fsub bfloat -0.0, %fabs
   %fmul = fmul bfloat %y, %fsub
@@ -854,80 +808,42 @@ define amdgpu_kernel void @fold_user_fneg_fabs_v2bf16(ptr addrspace(1) %out, <2
 ; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: fold_user_fneg_fabs_v2bf16:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_load_b32 s0, s[4:5], 0x8
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0x7fff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    v_mul_f32_e64 v0, s1, -4.0
-; GFX11-TRUE16-NEXT:    v_mul_f32_e64 v1, s0, -4.0
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v3, v5 :: v_dual_and_b32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_store_b32 v2, v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: fold_user_fneg_fabs_v2bf16:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b32 s0, s[4:5], 0x8
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s0, 0x7fff
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s0, 0x7fff
-; GFX11-FAKE16-NEXT:    v_mul_f32_e64 v0, s1, -4.0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_mul_f32_e64 v1, s0, -4.0
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, 0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v3, v5 :: v_dual_and_b32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_store_b32 v2, v0, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: fold_user_fneg_fabs_v2bf16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b32 s0, s[4:5], 0x8
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_and_b32 s1, s0, 0x7fff
+; GFX11-NEXT:    s_lshr_b32 s0, s0, 16
+; GFX11-NEXT:    s_lshl_b32 s1, s1, 16
+; GFX11-NEXT:    s_and_b32 s0, s0, 0x7fff
+; GFX11-NEXT:    v_mul_f32_e64 v0, s1, -4.0
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_mul_f32_e64 v1, s0, -4.0
+; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x0
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_bfe_u32 v3, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v3, v5 :: v_dual_and_b32 v0, 0xffff, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-NEXT:    s_endpgm
   %fabs = call <2 x bfloat> @llvm.fabs.v2bf16(<2 x bfloat> %in)
   %fneg.fabs = fsub <2 x bfloat> <bfloat -0.0, bfloat -0.0>, %fabs
   %mul = fmul <2 x bfloat> %fneg.fabs, <bfloat 4.0, bfloat 4.0>
@@ -1116,89 +1032,46 @@ define amdgpu_kernel void @s_fneg_multi_use_fabs_foldable_neg_v2bf16(ptr addrspa
 ; GFX9-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: s_fneg_multi_use_fabs_foldable_neg_v2bf16:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_load_b32 s6, s[4:5], 0x10
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s6
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s1, s6, 16
-; GFX11-TRUE16-NEXT:    s_and_b32 s0, s0, 0x7fff
-; GFX11-TRUE16-NEXT:    s_and_b32 s1, s1, 0x7fff
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_mul_f32_e64 v0, s0, -4.0
-; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s1, 16
-; GFX11-TRUE16-NEXT:    v_mul_f32_e64 v1, s0, -4.0
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v0
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
-; GFX11-TRUE16-NEXT:    s_and_b32 s4, s6, 0x7fff7fff
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_and_b32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    global_store_b32 v2, v3, s[0:1]
-; GFX11-TRUE16-NEXT:    global_store_b32 v2, v0, s[2:3]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: s_fneg_multi_use_fabs_foldable_neg_v2bf16:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b32 s6, s[4:5], 0x10
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_and_b32 s0, s6, 0x7fff
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s1, s6, 16
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 0x7fff
-; GFX11-FAKE16-NEXT:    v_mul_f32_e64 v0, s0, -4.0
-; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s1, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_mul_f32_e64 v1, s0, -4.0
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v0
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
-; GFX11-FAKE16-NEXT:    s_and_b32 s4, s6, 0x7fff7fff
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, 0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_and_b32 v0, 0xffff, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    global_store_b32 v2, v3, s[0:1]
-; GFX11-FAKE16-NEXT:    global_store_b32 v2, v0, s[2:3]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_fneg_multi_use_fabs_foldable_neg_v2bf16:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b32 s6, s[4:5], 0x10
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_and_b32 s0, s6, 0x7fff
+; GFX11-NEXT:    s_lshr_b32 s1, s6, 16
+; GFX11-NEXT:    s_lshl_b32 s0, s0, 16
+; GFX11-NEXT:    s_and_b32 s1, s1, 0x7fff
+; GFX11-NEXT:    v_mul_f32_e64 v0, s0, -4.0
+; GFX11-NEXT:    s_lshl_b32 s0, s1, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_mul_f32_e64 v1, s0, -4.0
+; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
+; GFX11-NEXT:    v_bfe_u32 v2, v0, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v0
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    v_bfe_u32 v3, v1, 16, 1
+; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v1
+; GFX11-NEXT:    s_and_b32 s4, s6, 0x7fff7fff
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, v3, v1
+; GFX11-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, v2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
+; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc_lo
+; GFX11-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_and_b32 v0, 0xffff, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    global_store_b32 v2, v3, s[0:1]
+; GFX11-NEXT:    global_store_b32 v2, v0, s[2:3]
+; GFX11-NEXT:    s_endpgm
   %fabs = call <2 x bfloat> @llvm.fabs.v2bf16(<2 x bfloat> %in)
   %fneg = fsub <2 x bfloat> <bfloat -0.0, bfloat -0.0>, %fabs
   %mul = fmul <2 x bfloat> %fneg, <bfloat 4.0, bfloat 4.0>
diff --git a/llvm/test/CodeGen/AMDGPU/frem.ll b/llvm/test/CodeGen/AMDGPU/frem.ll
index d3432daedadf8..6fb64a9678a69 100644
--- a/llvm/test/CodeGen/AMDGPU/frem.ll
+++ b/llvm/test/CodeGen/AMDGPU/frem.ll
@@ -2726,61 +2726,64 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i
 ; GFX1150-TRUE16-NEXT:    s_clause 0x1
 ; GFX1150-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 ; GFX1150-TRUE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX1150-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
+; GFX1150-TRUE16-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX1150-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1150-TRUE16-NEXT:    s_clause 0x1
-; GFX1150-TRUE16-NEXT:    global_load_b32 v2, v1, s[2:3]
-; GFX1150-TRUE16-NEXT:    global_load_b32 v3, v1, s[4:5] offset:16
+; GFX1150-TRUE16-NEXT:    global_load_b32 v3, v2, s[2:3]
+; GFX1150-TRUE16-NEXT:    global_load_b32 v4, v2, s[4:5] offset:16
 ; GFX1150-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.h
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v3.h
 ; GFX1150-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v4, v3.h
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(TRANS32_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v4, v4
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v4
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v4.h
+; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v5, -v3, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v5, v4
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v5, -v4, v0, v3 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v5, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v5, -v3, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v4, v5, v4
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v5, -v4, v0, v3 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v1, v5, v1
 ; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff800000, v4
-; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v4, v0
-; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1150-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v5.l
 ; GFX1150-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
-; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v5.l, v4.l
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v6.l, v1.l
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v4.l
 ; GFX1150-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX1150-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1150-TRUE16-NEXT:    v_fmac_f16_e32 v4.l, v0.l, v5.l
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v5, v3.l
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.l
-; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v5, v5
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v5
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v6, -v3, v0, v2 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_fmac_f16_e32 v5.l, v0.l, v6.l
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v3.l
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v6, v5
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v6, -v3, v0, v2 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v6, -v4, v0, v3 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v6, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v5, v6, v5
-; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff800000, v5
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v6, -v4, v0, v3 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v1, v6, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v5, v0
+; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1150-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v3.l
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX1150-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
+; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v4.l, v1.l
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v3.l, v2.l
 ; GFX1150-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX1150-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1150-TRUE16-NEXT:    v_fmac_f16_e32 v2.l, v0.l, v3.l
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_pack_b32_f16 v0, v2.l, v4.l
-; GFX1150-TRUE16-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_fmac_f16_e32 v3.l, v0.l, v4.l
+; GFX1150-TRUE16-NEXT:    v_pack_b32_f16 v0, v3.l, v5.l
+; GFX1150-TRUE16-NEXT:    global_store_b32 v2, v0, s[0:1]
 ; GFX1150-TRUE16-NEXT:    s_endpgm
 ;
 ; GFX1150-FAKE16-LABEL: frem_v2f16:
@@ -2851,61 +2854,64 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i
 ; GFX1200-TRUE16-NEXT:    s_clause 0x1
 ; GFX1200-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 ; GFX1200-TRUE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX1200-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
+; GFX1200-TRUE16-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX1200-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX1200-TRUE16-NEXT:    s_clause 0x1
-; GFX1200-TRUE16-NEXT:    global_load_b32 v2, v1, s[2:3]
-; GFX1200-TRUE16-NEXT:    global_load_b32 v3, v1, s[4:5] offset:16
+; GFX1200-TRUE16-NEXT:    global_load_b32 v3, v2, s[2:3]
+; GFX1200-TRUE16-NEXT:    global_load_b32 v4, v2, s[4:5] offset:16
 ; GFX1200-TRUE16-NEXT:    s_wait_loadcnt 0x1
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.h
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v3.h
 ; GFX1200-TRUE16-NEXT:    s_wait_loadcnt 0x0
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v4, v3.h
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(TRANS32_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v4, v4
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v4
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v4.h
+; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v5, -v3, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v5, v4
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v5, -v4, v0, v3 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v5, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v5, -v3, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v4, v5, v4
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v5, -v4, v0, v3 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v1, v5, v1
 ; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff800000, v4
-; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v4, v0
-; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1200-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v5.l
 ; GFX1200-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
-; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v5.l, v4.l
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v6.l, v1.l
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v4.l
 ; GFX1200-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX1200-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1200-TRUE16-NEXT:    v_fmac_f16_e32 v4.l, v0.l, v5.l
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v5, v3.l
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.l
-; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v5, v5
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v5
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v6, -v3, v0, v2 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_fmac_f16_e32 v5.l, v0.l, v6.l
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v3.l
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v6, v5
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v6, -v3, v0, v2 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v6, -v4, v0, v3 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v6, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v5, v6, v5
-; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v5, 0xff800000, v5
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v6, -v4, v0, v3 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v1, v6, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v5, v0
+; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1200-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v3.l
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX1200-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
+; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v4.l, v1.l
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v3.l, v2.l
 ; GFX1200-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX1200-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1200-TRUE16-NEXT:    v_fmac_f16_e32 v2.l, v0.l, v3.l
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_pack_b32_f16 v0, v2.l, v4.l
-; GFX1200-TRUE16-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_fmac_f16_e32 v3.l, v0.l, v4.l
+; GFX1200-TRUE16-NEXT:    v_pack_b32_f16 v0, v3.l, v5.l
+; GFX1200-TRUE16-NEXT:    global_store_b32 v2, v0, s[0:1]
 ; GFX1200-TRUE16-NEXT:    s_endpgm
 ;
 ; GFX1200-FAKE16-LABEL: frem_v2f16:
@@ -3474,57 +3480,54 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xff800000, v2
 ; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, v2, v7
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v1.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.h, v2
 ; GFX11-TRUE16-NEXT:    v_div_fixup_f16 v0.h, v0.h, v8.l, v6.l
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_trunc_f16_e32 v0.h, v0.h
 ; GFX11-TRUE16-NEXT:    v_fma_f16 v0.h, -v0.h, v8.l, v6.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_cvt_f32_f16_e32 v8, v7.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GFX11-TRUE16-NEXT:    v_pack_b32_f16 v2, v0.l, v0.h
 ; GFX11-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v1.l
+; GFX11-TRUE16-NEXT:    v_rcp_f32_e32 v8, v8
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v4
+; GFX11-TRUE16-NEXT:    v_fma_mix_f32 v6, -v3, v0, v1 op_sel_hi:[1,0,1]
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_fmac_f32_e32 v0, v6, v4
+; GFX11-TRUE16-NEXT:    v_fma_mix_f32 v6, -v3, v0, v1 op_sel_hi:[1,0,1]
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_fma_mix_f32 v8, -v6, v0, v7 op_sel_hi:[1,0,1]
-; GFX11-TRUE16-NEXT:    v_fmac_f32_e32 v0, v8, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_fma_mix_f32 v6, -v6, v0, v7 op_sel_hi:[1,0,1]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
 ; GFX11-TRUE16-NEXT:    v_mul_f32_e32 v4, v6, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cvt_f32_f16_e32 v8, v7.l
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xff800000, v4
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_rcp_f32_e32 v8, v8
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, v4, v0
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_cvt_f32_f16_e32 v6, v4.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v3.l, v1.l
-; GFX11-TRUE16-NEXT:    s_waitcnt_depctr 0xfff
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_mul_f32_e32 v6, v6, v8
 ; GFX11-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
-; GFX11-TRUE16-NEXT:    v_fma_mix_f32 v9, -v3, v6, v1 op_sel:[1,0,1] op_sel_hi:[1,0,1]
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_fma_mix_f32 v9, -v3, v6, v1 op_sel:[1,0,1] op_sel_hi:[1,0,1]
 ; GFX11-TRUE16-NEXT:    v_fma_f16 v0.l, -v0.l, v3.l, v1.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_fmac_f32_e32 v6, v9, v8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_fma_mix_f32 v1, -v3, v6, v1 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX11-TRUE16-NEXT:    v_mul_f32_e32 v1, v1, v8
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_mul_f32_e32 v1, v1, v8
 ; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
-; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, v1, v6
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, v1, v6
 ; GFX11-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.h, v1
-; GFX11-TRUE16-NEXT:    v_div_fixup_f16 v0.h, v0.h, v7.l, v4.l
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_div_fixup_f16 v0.h, v0.h, v7.l, v4.l
 ; GFX11-TRUE16-NEXT:    v_trunc_f16_e32 v0.h, v0.h
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_fma_f16 v0.h, -v0.h, v7.l, v4.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-TRUE16-NEXT:    v_pack_b32_f16 v1, v0.l, v0.h
 ; GFX11-TRUE16-NEXT:    global_store_b64 v5, v[1:2], s[0:1]
 ; GFX11-TRUE16-NEXT:    s_endpgm
@@ -3639,110 +3642,115 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i
 ; GFX1150-TRUE16-NEXT:    s_clause 0x1
 ; GFX1150-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 ; GFX1150-TRUE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX1150-TRUE16-NEXT:    v_mov_b32_e32 v5, 0
+; GFX1150-TRUE16-NEXT:    v_mov_b32_e32 v6, 0
 ; GFX1150-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1150-TRUE16-NEXT:    s_clause 0x1
-; GFX1150-TRUE16-NEXT:    global_load_b64 v[1:2], v5, s[2:3]
-; GFX1150-TRUE16-NEXT:    global_load_b64 v[3:4], v5, s[4:5] offset:32
+; GFX1150-TRUE16-NEXT:    global_load_b64 v[2:3], v6, s[2:3]
+; GFX1150-TRUE16-NEXT:    global_load_b64 v[4:5], v6, s[4:5] offset:32
 ; GFX1150-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v1.h
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.h
 ; GFX1150-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v6, v3.h
-; GFX1150-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v3.l
-; GFX1150-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v1.l
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(TRANS32_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v6, v6
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v6
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v4.h
+; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v4
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v7, -v3, v0, v1 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v7, v6
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v7, -v4, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v7, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v7, -v3, v0, v1 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v6, v7, v6
-; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff800000, v6
-; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v6, v0
-; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v7, -v4, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v1, v7, v1
+; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v2
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1150-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v7.l
 ; GFX1150-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
-; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v7.l, v6.l
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v8.l, v1.l
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v4.l
 ; GFX1150-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX1150-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1150-TRUE16-NEXT:    v_fmac_f16_e32 v6.l, v0.l, v7.l
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v7, v3.l
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v1.l
-; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v7, v7
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v7
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v10, -v8, v0, v9 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_fmac_f16_e32 v7.l, v0.l, v8.l
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.l
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v10, v7
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v8, -v8, v0, v9 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v8, -v4, v0, v2 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v8, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v7, v8, v7
-; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff800000, v7
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v8, -v4, v0, v2 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v1, v8, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v7, v0
+; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v5.h
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GFX1150-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
+; GFX1150-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v2.l
+; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v3.l, v1.l
+; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v4.l, v0.h
 ; GFX1150-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX1150-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1150-TRUE16-NEXT:    v_fma_f16 v0.l, v0.l, v3.l, v1.l
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v3, v4.h
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1150-TRUE16-NEXT:    v_pack_b32_f16 v1, v0.l, v6.l
-; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v3, v3
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.h
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v3
+; GFX1150-TRUE16-NEXT:    v_fma_f16 v0.l, v0.l, v4.l, v2.l
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_pack_b32_f16 v2, v0.l, v7.l
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v3.h
+; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_2)
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v6, -v4, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v6, v3
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v4, -v5, v0, v3 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v4, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v6, -v4, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v3, v6, v3
-; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff800000, v3
-; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v3, v0
-; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v4, -v5, v0, v3 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v1, v4, v1
+; GFX1150-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1150-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v4.l
 ; GFX1150-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
-; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v6.l, v3.l
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v7.l, v1.l
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v5.l
 ; GFX1150-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX1150-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1150-TRUE16-NEXT:    v_fmac_f16_e32 v3.l, v0.l, v6.l
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v6, v4.l
-; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.l
-; GFX1150-TRUE16-NEXT:    v_rcp_f32_e32 v6, v6
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v6
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v7, -v4, v0, v2 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_fmac_f16_e32 v4.l, v0.l, v7.l
+; GFX1150-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v3.l
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v7, v6
-; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v7, -v4, v0, v2 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v7, -v5, v0, v3 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_fmac_f32_e32 v0, v7, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v6, v7, v6
-; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff800000, v6
+; GFX1150-TRUE16-NEXT:    v_fma_mix_f32 v7, -v5, v0, v3 op_sel_hi:[1,0,1]
+; GFX1150-TRUE16-NEXT:    v_mul_f32_e32 v1, v7, v1
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v6, v0
+; GFX1150-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1150-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
 ; GFX1150-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
+; GFX1150-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v3.l
+; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v5.l, v0.h
 ; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v4.l, v2.l
 ; GFX1150-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX1150-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1150-TRUE16-NEXT:    v_fma_f16 v0.l, v0.l, v4.l, v2.l
-; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX1150-TRUE16-NEXT:    v_pack_b32_f16 v2, v0.l, v3.l
-; GFX1150-TRUE16-NEXT:    global_store_b64 v5, v[1:2], s[0:1]
+; GFX1150-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1150-TRUE16-NEXT:    v_fma_f16 v0.l, v0.l, v5.l, v3.l
+; GFX1150-TRUE16-NEXT:    v_pack_b32_f16 v3, v0.l, v4.l
+; GFX1150-TRUE16-NEXT:    global_store_b64 v6, v[2:3], s[0:1]
 ; GFX1150-TRUE16-NEXT:    s_endpgm
 ;
 ; GFX1150-FAKE16-LABEL: frem_v4f16:
@@ -3860,110 +3868,115 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i
 ; GFX1200-TRUE16-NEXT:    s_clause 0x1
 ; GFX1200-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
 ; GFX1200-TRUE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
-; GFX1200-TRUE16-NEXT:    v_mov_b32_e32 v5, 0
+; GFX1200-TRUE16-NEXT:    v_mov_b32_e32 v6, 0
 ; GFX1200-TRUE16-NEXT:    s_wait_kmcnt 0x0
 ; GFX1200-TRUE16-NEXT:    s_clause 0x1
-; GFX1200-TRUE16-NEXT:    global_load_b64 v[1:2], v5, s[2:3]
-; GFX1200-TRUE16-NEXT:    global_load_b64 v[3:4], v5, s[4:5] offset:32
+; GFX1200-TRUE16-NEXT:    global_load_b64 v[2:3], v6, s[2:3]
+; GFX1200-TRUE16-NEXT:    global_load_b64 v[4:5], v6, s[4:5] offset:32
 ; GFX1200-TRUE16-NEXT:    s_wait_loadcnt 0x1
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v1.h
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.h
 ; GFX1200-TRUE16-NEXT:    s_wait_loadcnt 0x0
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v6, v3.h
-; GFX1200-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v3.l
-; GFX1200-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v1.l
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(TRANS32_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v6, v6
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v6
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v4.h
+; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v4
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v7, -v3, v0, v1 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v7, v6
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v7, -v4, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v7, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v7, -v3, v0, v1 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v6, v7, v6
-; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff800000, v6
-; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v6, v0
-; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v7, -v4, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v1, v7, v1
+; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v2
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1200-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v7.l
 ; GFX1200-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
-; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v7.l, v6.l
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v8.l, v1.l
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v4.l
 ; GFX1200-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX1200-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1200-TRUE16-NEXT:    v_fmac_f16_e32 v6.l, v0.l, v7.l
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v7, v3.l
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v1.l
-; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v7, v7
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v7
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v10, -v8, v0, v9 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_fmac_f16_e32 v7.l, v0.l, v8.l
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.l
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v10, v7
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v8, -v8, v0, v9 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v8, -v4, v0, v2 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v8, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v7, v8, v7
-; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v7, 0xff800000, v7
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v8, -v4, v0, v2 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v1, v8, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v7, v0
+; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v5.h
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GFX1200-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
+; GFX1200-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v2.l
+; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v3.l, v1.l
+; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v4.l, v0.h
 ; GFX1200-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX1200-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1200-TRUE16-NEXT:    v_fma_f16 v0.l, v0.l, v3.l, v1.l
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v3, v4.h
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1200-TRUE16-NEXT:    v_pack_b32_f16 v1, v0.l, v6.l
-; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v3, v3
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.h
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v3
+; GFX1200-TRUE16-NEXT:    v_fma_f16 v0.l, v0.l, v4.l, v2.l
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_pack_b32_f16 v2, v0.l, v7.l
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v3.h
+; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v5
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_2)
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v6, -v4, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v6, v3
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v4, -v5, v0, v3 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v4, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v6, -v4, v0, v2 op_sel:[1,0,1] op_sel_hi:[1,0,1]
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v3, v6, v3
-; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v3, 0xff800000, v3
-; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v3, v0
-; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v4, -v5, v0, v3 op_sel:[1,0,1] op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v1, v4, v1
+; GFX1200-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1200-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v4.l
 ; GFX1200-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
-; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v6.l, v3.l
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v7.l, v1.l
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v1, v5.l
 ; GFX1200-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX1200-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1200-TRUE16-NEXT:    v_fmac_f16_e32 v3.l, v0.l, v6.l
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v6, v4.l
-; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v2.l
-; GFX1200-TRUE16-NEXT:    v_rcp_f32_e32 v6, v6
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v6
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v7, -v4, v0, v2 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_fmac_f16_e32 v4.l, v0.l, v7.l
+; GFX1200-TRUE16-NEXT:    v_cvt_f32_f16_e32 v0, v3.l
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(TRANS32_DEP_1) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v0, v0, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v7, v6
-; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v7, -v4, v0, v2 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v7, -v5, v0, v3 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_fmac_f32_e32 v0, v7, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v6, v7, v6
-; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v6, 0xff800000, v6
+; GFX1200-TRUE16-NEXT:    v_fma_mix_f32 v7, -v5, v0, v3 op_sel_hi:[1,0,1]
+; GFX1200-TRUE16-NEXT:    v_mul_f32_e32 v1, v7, v1
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v6, v0
+; GFX1200-TRUE16-NEXT:    v_and_b32_e32 v1, 0xff800000, v1
+; GFX1200-TRUE16-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
 ; GFX1200-TRUE16-NEXT:    v_cvt_f16_f32_e32 v0.l, v0
+; GFX1200-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v3.l
+; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v5.l, v0.h
 ; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_div_fixup_f16 v0.l, v0.l, v4.l, v2.l
 ; GFX1200-TRUE16-NEXT:    v_trunc_f16_e32 v0.l, v0.l
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX1200-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x8000, v0
-; GFX1200-TRUE16-NEXT:    v_fma_f16 v0.l, v0.l, v4.l, v2.l
-; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX1200-TRUE16-NEXT:    v_pack_b32_f16 v2, v0.l, v3.l
-; GFX1200-TRUE16-NEXT:    global_store_b64 v5, v[1:2], s[0:1]
+; GFX1200-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-TRUE16-NEXT:    v_fma_f16 v0.l, v0.l, v5.l, v3.l
+; GFX1200-TRUE16-NEXT:    v_pack_b32_f16 v3, v0.l, v4.l
+; GFX1200-TRUE16-NEXT:    global_store_b64 v6, v[2:3], s[0:1]
 ; GFX1200-TRUE16-NEXT:    s_endpgm
 ;
 ; GFX1200-FAKE16-LABEL: frem_v4f16:
diff --git a/llvm/test/CodeGen/AMDGPU/function-returns.ll b/llvm/test/CodeGen/AMDGPU/function-returns.ll
index 08515fcc08e5e..658c45c5b2886 100644
--- a/llvm/test/CodeGen/AMDGPU/function-returns.ll
+++ b/llvm/test/CodeGen/AMDGPU/function-returns.ll
@@ -1177,59 +1177,32 @@ define <16 x i8> @v16i8_func_void() #0 {
 ; GFX789-NEXT:    v_mov_b32_e32 v3, v18
 ; GFX789-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: v16i8_func_void:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, 0x31016000
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, -1
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    buffer_load_b128 v[16:19], off, s[0:3], 0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 24, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v16.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v17.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v18.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v19.l
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: v16i8_func_void:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s3, 0x31016000
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, -1
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    buffer_load_b128 v[0:3], off, s[0:3], 0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 24, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v3
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v16
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, v3 :: v_dual_mov_b32 v3, v18
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, v17
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: v16i8_func_void:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s3, 0x31016000
+; GFX11-NEXT:    s_mov_b32 s2, -1
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    buffer_load_b128 v[0:3], off, s[0:3], 0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 8, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 24, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v3
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v16
+; GFX11-NEXT:    v_mov_b32_e32 v8, v2
+; GFX11-NEXT:    v_dual_mov_b32 v12, v3 :: v_dual_mov_b32 v3, v18
+; GFX11-NEXT:    v_mov_b32_e32 v2, v17
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %ptr = load volatile ptr addrspace(1), ptr addrspace(4) poison
   %val = load <16 x i8>, ptr addrspace(1) %ptr
   ret <16 x i8> %val
diff --git a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
index e40917d4307fb..dd7f18357f965 100644
--- a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
@@ -3616,76 +3616,40 @@ define amdgpu_gfx void @test_call_external_void_func_v5i8() #0 {
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: test_call_external_void_func_v5i8:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s33
-; GFX11-TRUE16-NEXT:    s_mov_b32 s33, s32
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, external_void_func_v5i8 at abs32@hi
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, external_void_func_v5i8 at abs32@lo
-; GFX11-TRUE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-TRUE16-NEXT:    global_load_b64 v[3:4], v[0:1], off
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v5.l
-; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s32, s33
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s33, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: test_call_external_void_func_v5i8:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, s33
-; GFX11-FAKE16-NEXT:    s_mov_b32 s33, s32
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, external_void_func_v5i8 at abs32@hi
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, external_void_func_v5i8 at abs32@lo
-; GFX11-FAKE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-FAKE16-NEXT:    global_load_b64 v[5:6], v[0:1], off
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[5:6]
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v5
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, v5
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v6
-; GFX11-FAKE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s32, s33
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s33, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: test_call_external_void_func_v5i8:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_mov_b32 s0, s33
+; GFX11-NEXT:    s_mov_b32 s33, s32
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_mov_b32_e32 v0, 0
+; GFX11-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v5i8 at abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v5i8 at abs32@lo
+; GFX11-NEXT:    s_add_i32 s32, s32, 16
+; GFX11-NEXT:    global_load_b64 v[5:6], v[0:1], off
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_lshrrev_b64 v[3:4], 24, v[5:6]
+; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 8, v5
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v5
+; GFX11-NEXT:    v_mov_b32_e32 v0, v5
+; GFX11-NEXT:    v_mov_b32_e32 v4, v6
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX11-NEXT:    s_mov_b32 s32, s33
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    s_mov_b32 s33, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_v5i8:
 ; GFX10-SCRATCH:       ; %bb.0:
@@ -3807,80 +3771,42 @@ define amdgpu_gfx void @test_call_external_void_func_v8i8() #0 {
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: test_call_external_void_func_v8i8:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s33
-; GFX11-TRUE16-NEXT:    s_mov_b32 s33, s32
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, external_void_func_v8i8 at abs32@hi
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, external_void_func_v8i8 at abs32@lo
-; GFX11-TRUE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-TRUE16-NEXT:    global_load_b64 v[8:9], v[0:1], off
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v9.l
-; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s32, s33
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s33, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: test_call_external_void_func_v8i8:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, s33
-; GFX11-FAKE16-NEXT:    s_mov_b32 s33, s32
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, external_void_func_v8i8 at abs32@hi
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, external_void_func_v8i8 at abs32@lo
-; GFX11-FAKE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-FAKE16-NEXT:    global_load_b64 v[0:1], v[0:1], off
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 8, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v8
-; GFX11-FAKE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s32, s33
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s33, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: test_call_external_void_func_v8i8:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_mov_b32 s0, s33
+; GFX11-NEXT:    s_mov_b32 s33, s32
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_mov_b32_e32 v0, 0
+; GFX11-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v8i8 at abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v8i8 at abs32@lo
+; GFX11-NEXT:    s_add_i32 s32, s32, 16
+; GFX11-NEXT:    global_load_b64 v[0:1], v[0:1], off
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v8, 8, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v8
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX11-NEXT:    s_mov_b32 s32, s33
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    s_mov_b32 s33, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_v8i8:
 ; GFX10-SCRATCH:       ; %bb.0:
@@ -4070,137 +3996,71 @@ define amdgpu_gfx void @test_call_external_void_func_v32i8() #0 {
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: test_call_external_void_func_v32i8:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s33
-; GFX11-TRUE16-NEXT:    s_mov_b32 s33, s32
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, external_void_func_v32i8 at abs32@hi
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, external_void_func_v32i8 at abs32@lo
-; GFX11-TRUE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-TRUE16-NEXT:    global_load_b128 v[32:35], v[0:1], off
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, 16
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-TRUE16-NEXT:    global_load_b128 v[36:39], v[0:1], off
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v34
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 24, v34
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v35
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 24, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 24, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v32.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v33.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v34.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v35.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v36.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v37.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v38.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v39.l
-; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s32, s33
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s33, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: test_call_external_void_func_v32i8:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, s33
-; GFX11-FAKE16-NEXT:    s_mov_b32 s33, s32
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v4, 16
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v5, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, external_void_func_v32i8 at abs32@hi
-; GFX11-FAKE16-NEXT:    global_load_b128 v[0:3], v[0:1], off
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, external_void_func_v32i8 at abs32@lo
-; GFX11-FAKE16-NEXT:    global_load_b128 v[16:19], v[4:5], off
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-FAKE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 24, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 8, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 24, v16
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v17
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 24, v18
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v19
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v35
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v2
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v12, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v20, v17
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v24, v18
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v28, v19
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v36 :: v_dual_mov_b32 v3, v37
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, v32 :: v_dual_mov_b32 v18, v33
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v19, v34
-; GFX11-FAKE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s32, s33
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s33, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: test_call_external_void_func_v32i8:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_mov_b32 s0, s33
+; GFX11-NEXT:    s_mov_b32 s33, s32
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_mov_b32_e32 v0, 0
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v4, 16
+; GFX11-NEXT:    v_mov_b32_e32 v5, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v32i8 at abs32@hi
+; GFX11-NEXT:    global_load_b128 v[0:3], v[0:1], off
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v32i8 at abs32@lo
+; GFX11-NEXT:    global_load_b128 v[16:19], v[4:5], off
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    s_add_i32 s32, s32, 16
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX11-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v35, 8, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v36, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v37, 24, v0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 8, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 24, v16
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v21, 8, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v23, 24, v17
+; GFX11-NEXT:    v_lshrrev_b32_e32 v25, 8, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 16, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v27, 24, v18
+; GFX11-NEXT:    v_lshrrev_b32_e32 v29, 8, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v19
+; GFX11-NEXT:    v_lshrrev_b32_e32 v31, 24, v19
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v35
+; GFX11-NEXT:    v_mov_b32_e32 v8, v2
+; GFX11-NEXT:    v_mov_b32_e32 v12, v3
+; GFX11-NEXT:    v_mov_b32_e32 v20, v17
+; GFX11-NEXT:    v_mov_b32_e32 v24, v18
+; GFX11-NEXT:    v_mov_b32_e32 v28, v19
+; GFX11-NEXT:    v_dual_mov_b32 v2, v36 :: v_dual_mov_b32 v3, v37
+; GFX11-NEXT:    v_dual_mov_b32 v17, v32 :: v_dual_mov_b32 v18, v33
+; GFX11-NEXT:    v_mov_b32_e32 v19, v34
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX11-NEXT:    s_mov_b32 s32, s33
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    s_mov_b32 s33, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_v32i8:
 ; GFX10-SCRATCH:       ; %bb.0:
@@ -5285,20 +5145,19 @@ define amdgpu_gfx void @test_call_external_void_func_v5i8_ret() #0 {
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s1, external_void_func_v5i8_ret at abs32@hi
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, external_void_func_v5i8_ret at abs32@lo
 ; GFX11-TRUE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-TRUE16-NEXT:    global_load_b64 v[3:4], v[40:41], off
+; GFX11-TRUE16-NEXT:    global_load_b64 v[5:6], v[40:41], off
 ; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s30, 0
 ; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s31, 1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[5:6], 24, v[3:4]
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v3.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v5.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[5:6]
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v5
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v5
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v5
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v4, v6
 ; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v0.h, 8, v1.l
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
 ; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.l, 8, v3.l
 ; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.l
 ; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v0.l, v0.h
@@ -5568,18 +5427,17 @@ define amdgpu_gfx void @test_call_external_void_func_v8i8_ret() #0 {
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s1, external_void_func_v8i8_ret at abs32@hi
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, external_void_func_v8i8_ret at abs32@lo
 ; GFX11-TRUE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-TRUE16-NEXT:    global_load_b64 v[8:9], v[40:41], off
+; GFX11-TRUE16-NEXT:    global_load_b64 v[0:1], v[40:41], off
 ; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s30, 0
 ; GFX11-TRUE16-NEXT:    v_writelane_b32 v42, s31, 1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v8
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v9
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v9
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v8.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v9.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 8, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v8
 ; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v0.h, 8, v5.l
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
@@ -6009,46 +5867,47 @@ define amdgpu_gfx void @test_call_external_void_func_v32i8_ret() #0 {
 ; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v43, 0
 ; GFX11-TRUE16-NEXT:    v_writelane_b32 v44, s0, 2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s1, external_void_func_v3i8_ret at abs32@hi
-; GFX11-TRUE16-NEXT:    global_load_b128 v[32:35], v[40:41], off
+; GFX11-TRUE16-NEXT:    global_load_b128 v[0:3], v[40:41], off
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, external_void_func_v3i8_ret at abs32@lo
-; GFX11-TRUE16-NEXT:    global_load_b128 v[36:39], v[42:43], off
+; GFX11-TRUE16-NEXT:    global_load_b128 v[16:19], v[42:43], off
 ; GFX11-TRUE16-NEXT:    v_writelane_b32 v44, s30, 0
 ; GFX11-TRUE16-NEXT:    s_add_i32 s32, s32, 32
 ; GFX11-TRUE16-NEXT:    v_writelane_b32 v44, s31, 1
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v32
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v33
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v34
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v34
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 24, v34
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v35
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v35
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v0
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 24, v0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 24, v36
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v37
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 24, v38
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v39
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v39
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v32.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v33.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v34.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v35.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v36.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v37.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v38.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v39.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 8, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 24, v16
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 24, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 8, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 24, v17
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 8, v18
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v18
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 24, v18
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 8, v19
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v19
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v19
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v35
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, v2
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v12, v3
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v20, v17
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v24, v18
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v28, v19
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v36 :: v_dual_mov_b32 v3, v37
+; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v17, v32 :: v_dual_mov_b32 v18, v33
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v19, v34
 ; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 ; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v0.h, 8, v13.l
 ; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v12.l
@@ -9825,99 +9684,52 @@ define amdgpu_gfx void @test_call_external_void_func_v16i8() #0 {
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: test_call_external_void_func_v16i8:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s33
-; GFX11-TRUE16-NEXT:    s_mov_b32 s33, s32
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-TRUE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_load_b128 v[16:19], v0, s[0:1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, external_void_func_v16i8 at abs32@hi
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, external_void_func_v16i8 at abs32@lo
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 8, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 24, v16
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v17
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 24, v18
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v19
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v19
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v16.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v17.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v18.l
-; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v19.l
-; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s32, s33
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s33, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: test_call_external_void_func_v16i8:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, s33
-; GFX11-FAKE16-NEXT:    s_mov_b32 s33, s32
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-FAKE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_load_b128 v[0:3], v0, s[0:1]
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, external_void_func_v16i8 at abs32@hi
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, external_void_func_v16i8 at abs32@lo
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 24, v0
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 24, v2
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 8, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
-; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 24, v3
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v4, v1
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, v2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, v3 :: v_dual_mov_b32 v1, v16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, v17 :: v_dual_mov_b32 v3, v18
-; GFX11-FAKE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s32, s33
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s33, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: test_call_external_void_func_v16i8:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_mov_b32 s0, s33
+; GFX11-NEXT:    s_mov_b32 s33, s32
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
+; GFX11-NEXT:    v_mov_b32_e32 v0, 0
+; GFX11-NEXT:    s_add_i32 s32, s32, 16
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    global_load_b128 v[0:3], v0, s[0:1]
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_v16i8 at abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_v16i8 at abs32@lo
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_lshrrev_b32_e32 v16, 8, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v18, 24, v0
+; GFX11-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v7, 24, v1
+; GFX11-NEXT:    v_lshrrev_b32_e32 v9, 8, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v11, 24, v2
+; GFX11-NEXT:    v_lshrrev_b32_e32 v13, 8, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
+; GFX11-NEXT:    v_lshrrev_b32_e32 v15, 24, v3
+; GFX11-NEXT:    v_mov_b32_e32 v4, v1
+; GFX11-NEXT:    v_mov_b32_e32 v8, v2
+; GFX11-NEXT:    v_dual_mov_b32 v12, v3 :: v_dual_mov_b32 v1, v16
+; GFX11-NEXT:    v_dual_mov_b32 v2, v17 :: v_dual_mov_b32 v3, v18
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX11-NEXT:    s_mov_b32 s32, s33
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    s_mov_b32 s33, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_v16i8:
 ; GFX10-SCRATCH:       ; %bb.0:
@@ -18221,61 +18033,32 @@ define amdgpu_gfx void @test_call_external_void_func_bf16_inreg(i16 inreg %arg)
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
-; GFX11-TRUE16-LABEL: test_call_external_void_func_bf16_inreg:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s33
-; GFX11-TRUE16-NEXT:    s_mov_b32 s33, s32
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s0, 3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, external_void_func_bf16 at abs32@hi
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, external_void_func_bf16 at abs32@lo
-; GFX11-TRUE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s4, 0
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s30, 1
-; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s31, 2
-; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v40, 2
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v40, 1
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s4, v40, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s32, s33
-; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v40, 3
-; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s33, s0
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11-FAKE16-LABEL: test_call_external_void_func_bf16_inreg:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, s33
-; GFX11-FAKE16-NEXT:    s_mov_b32 s33, s32
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s0, 2
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, external_void_func_bf16 at abs32@hi
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, external_void_func_bf16 at abs32@lo
-; GFX11-FAKE16-NEXT:    s_add_i32 s32, s32, 16
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s30, 0
-; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s31, 1
-; GFX11-FAKE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v40, 1
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v40, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s32, s33
-; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v40, 2
-; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s1, -1
-; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
-; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s33, s0
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+; GFX11-LABEL: test_call_external_void_func_bf16_inreg:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_mov_b32 s0, s33
+; GFX11-NEXT:    s_mov_b32 s33, s32
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX11-NEXT:    scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    v_writelane_b32 v40, s0, 2
+; GFX11-NEXT:    s_mov_b32 s1, external_void_func_bf16 at abs32@hi
+; GFX11-NEXT:    s_mov_b32 s0, external_void_func_bf16 at abs32@lo
+; GFX11-NEXT:    s_add_i32 s32, s32, 16
+; GFX11-NEXT:    v_writelane_b32 v40, s30, 0
+; GFX11-NEXT:    v_writelane_b32 v40, s31, 1
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_readlane_b32 s31, v40, 1
+; GFX11-NEXT:    v_readlane_b32 s30, v40, 0
+; GFX11-NEXT:    s_mov_b32 s32, s33
+; GFX11-NEXT:    v_readlane_b32 s0, v40, 2
+; GFX11-NEXT:    s_or_saveexec_b32 s1, -1
+; GFX11-NEXT:    scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT:    s_mov_b32 exec_lo, s1
+; GFX11-NEXT:    s_mov_b32 s33, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-SCRATCH-LABEL: test_call_external_void_func_bf16_inreg:
 ; GFX10-SCRATCH:       ; %bb.0:
diff --git a/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll b/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll
index 8196ffae2ca27..9389f1614721f 100644
--- a/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll
+++ b/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll
@@ -1,6 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs %s -o - | FileCheck %s -check-prefixes=GFX11,GFX11-TRUE16
-; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs %s -o - | FileCheck %s -check-prefixes=GFX11,GFX11-FAKE16
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs %s -o - | FileCheck %s -check-prefixes=GFX11
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs %s -o - | FileCheck %s -check-prefixes=GFX11
 
 declare i32 @llvm.amdgcn.workitem.id.x()
 
@@ -51,313 +51,160 @@ bb:
 
 ; FIXME: This generates "instid1(/* invalid instid value */)".
 define amdgpu_kernel void @f2(i32 %arg, i32 %arg1, i32 %arg2, i1 %arg3, i32 %arg4, i1 %arg5, ptr %arg6, i32 %arg7, i32 %arg8, i32 %arg9, i32 %arg10, i1 %arg11) {
-; GFX11-TRUE16-LABEL: f2:
-; GFX11-TRUE16:       ; %bb.0: ; %bb
-; GFX11-TRUE16-NEXT:    s_mov_b64 s[16:17], s[4:5]
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v31, v0
-; GFX11-TRUE16-NEXT:    s_load_b32 s19, s[16:17], 0x24
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s13
-; GFX11-TRUE16-NEXT:    s_mov_b64 s[10:11], s[6:7]
-; GFX11-TRUE16-NEXT:    s_mov_b64 s[6:7], s[2:3]
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v31
-; GFX11-TRUE16-NEXT:    s_mov_b64 s[4:5], s[0:1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s20, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, -1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, exec_lo
-; GFX11-TRUE16-NEXT:    s_mov_b32 s32, 0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_mul_lo_u32 v0, s19, v0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_cmpx_eq_u32_e32 0, v0
-; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB2_13
-; GFX11-TRUE16-NEXT:  ; %bb.1: ; %bb14
-; GFX11-TRUE16-NEXT:    s_load_b128 s[20:23], s[16:17], 0x2c
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, 0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_bitcmp1_b32 s21, 0
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s24, -1, 0
-; GFX11-TRUE16-NEXT:    s_bitcmp0_b32 s21, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB2_3
-; GFX11-TRUE16-NEXT:  ; %bb.2: ; %bb15
-; GFX11-TRUE16-NEXT:    s_add_u32 s8, s16, 0x58
-; GFX11-TRUE16-NEXT:    s_addc_u32 s9, s17, 0
-; GFX11-TRUE16-NEXT:    s_getpc_b64 s[0:1]
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, f0 at gotpcrel32@lo+4
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, f0 at gotpcrel32@hi+12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s14
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s21, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s15
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s21
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, -1
-; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB2_4
-; GFX11-TRUE16-NEXT:    s_branch .LBB2_12
-; GFX11-TRUE16-NEXT:  .LBB2_3:
-; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-TRUE16-NEXT:  .LBB2_4: ; %bb16
-; GFX11-TRUE16-NEXT:    s_load_b32 s1, s[16:17], 0x54
-; GFX11-TRUE16-NEXT:    s_bitcmp1_b32 s23, 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, -1
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s2, -1, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s9, s23, 1
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_bitcmp1_b32 s1, 0
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s1, -1, 0
-; GFX11-TRUE16-NEXT:    s_cmp_eq_u32 s9, 0
-; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB2_8
-; GFX11-TRUE16-NEXT:  ; %bb.5: ; %bb18.preheader
-; GFX11-TRUE16-NEXT:    s_load_b128 s[28:31], s[16:17], 0x44
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_mul_hi_u32 s8, s29, s28
-; GFX11-TRUE16-NEXT:    s_mul_i32 s9, s29, s28
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_alignbit_b32 v0, s8, s9, 1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s9, 0
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s8, 1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s8, s8, s30
-; GFX11-TRUE16-NEXT:    s_mul_i32 s8, s8, s22
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_mul_i32 s8, s8, s20
-; GFX11-TRUE16-NEXT:    s_or_b32 s8, s19, s8
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_lshl_b64 s[20:21], s[8:9], 1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, s9
-; GFX11-TRUE16-NEXT:    global_load_u16 v1, v0, s[20:21]
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s24
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v1
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v1, 0, 1, vcc_lo
-; GFX11-TRUE16-NEXT:    s_mov_b32 vcc_lo, 0
-; GFX11-TRUE16-NEXT:    .p2align 6
-; GFX11-TRUE16-NEXT:  .LBB2_6: ; %bb18
-; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, 0xffff, s8
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v0
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s8, 0
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, -1, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s1, s8
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, exec_lo
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s19, v2
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s8, s19, s13
-; GFX11-TRUE16-NEXT:    s_and_b32 s13, 0xffff, s9
-; GFX11-TRUE16-NEXT:    s_and_b32 s8, s8, 1
-; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s13, 0
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s13, -1, 0
-; GFX11-TRUE16-NEXT:    s_and_b32 s20, s2, exec_lo
-; GFX11-TRUE16-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s13
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s19, v2
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s13, s19, s13
-; GFX11-TRUE16-NEXT:    s_bitcmp1_b32 s13, 0
-; GFX11-TRUE16-NEXT:    s_cselect_b32 s13, 0x100, 0
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_or_b32 s9, s13, s9
-; GFX11-TRUE16-NEXT:    s_cbranch_vccz .LBB2_6
-; GFX11-TRUE16-NEXT:  ; %bb.7: ; %Flow
-; GFX11-TRUE16-NEXT:    s_mov_b32 s8, 0
-; GFX11-TRUE16-NEXT:  .LBB2_8: ; %Flow12
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_and_b32 vcc_lo, exec_lo, s8
-; GFX11-TRUE16-NEXT:    s_cbranch_vccz .LBB2_12
-; GFX11-TRUE16-NEXT:  ; %bb.9:
-; GFX11-TRUE16-NEXT:    s_xor_b32 s1, s1, -1
-; GFX11-TRUE16-NEXT:  .LBB2_10: ; %bb17
-; GFX11-TRUE16-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_and_b32 vcc_lo, exec_lo, s1
-; GFX11-TRUE16-NEXT:    s_cbranch_vccz .LBB2_10
-; GFX11-TRUE16-NEXT:  ; %bb.11: ; %Flow6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s18, -1
-; GFX11-TRUE16-NEXT:  .LBB2_12: ; %Flow11
-; GFX11-TRUE16-NEXT:    s_and_b32 s20, s0, exec_lo
-; GFX11-TRUE16-NEXT:    s_or_not1_b32 s0, s18, exec_lo
-; GFX11-TRUE16-NEXT:  .LBB2_13: ; %Flow9
-; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s3
-; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s3, s0
-; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB2_15
-; GFX11-TRUE16-NEXT:  ; %bb.14: ; %bb43
-; GFX11-TRUE16-NEXT:    s_add_u32 s8, s16, 0x58
-; GFX11-TRUE16-NEXT:    s_addc_u32 s9, s17, 0
-; GFX11-TRUE16-NEXT:    s_getpc_b64 s[0:1]
-; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, f0 at gotpcrel32@lo+4
-; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, f0 at gotpcrel32@hi+12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s14
-; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s15
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT:    s_or_b32 s20, s20, exec_lo
-; GFX11-TRUE16-NEXT:  .LBB2_15: ; %Flow14
-; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s3
-; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, s20
-; GFX11-TRUE16-NEXT:  ; %bb.16: ; %UnifiedUnreachableBlock
-; GFX11-TRUE16-NEXT:    ; divergent unreachable
-; GFX11-TRUE16-NEXT:  ; %bb.17: ; %UnifiedReturnBlock
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: f2:
-; GFX11-FAKE16:       ; %bb.0: ; %bb
-; GFX11-FAKE16-NEXT:    s_mov_b64 s[16:17], s[4:5]
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v31, v0
-; GFX11-FAKE16-NEXT:    s_load_b32 s19, s[16:17], 0x24
-; GFX11-FAKE16-NEXT:    s_mov_b32 s12, s13
-; GFX11-FAKE16-NEXT:    s_mov_b64 s[10:11], s[6:7]
-; GFX11-FAKE16-NEXT:    s_mov_b64 s[6:7], s[2:3]
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v31
-; GFX11-FAKE16-NEXT:    s_mov_b64 s[4:5], s[0:1]
-; GFX11-FAKE16-NEXT:    s_mov_b32 s20, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, -1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s3, exec_lo
-; GFX11-FAKE16-NEXT:    s_mov_b32 s32, 0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_mul_lo_u32 v0, s19, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_cmpx_eq_u32_e32 0, v0
-; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB2_13
-; GFX11-FAKE16-NEXT:  ; %bb.1: ; %bb14
-; GFX11-FAKE16-NEXT:    s_load_b128 s[20:23], s[16:17], 0x2c
-; GFX11-FAKE16-NEXT:    s_mov_b32 s18, 0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_bitcmp1_b32 s21, 0
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s24, -1, 0
-; GFX11-FAKE16-NEXT:    s_bitcmp0_b32 s21, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB2_3
-; GFX11-FAKE16-NEXT:  ; %bb.2: ; %bb15
-; GFX11-FAKE16-NEXT:    s_add_u32 s8, s16, 0x58
-; GFX11-FAKE16-NEXT:    s_addc_u32 s9, s17, 0
-; GFX11-FAKE16-NEXT:    s_getpc_b64 s[0:1]
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, f0 at gotpcrel32@lo+4
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, f0 at gotpcrel32@hi+12
-; GFX11-FAKE16-NEXT:    s_mov_b32 s13, s14
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s21, s14
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, s15
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, s21
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, -1
-; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB2_4
-; GFX11-FAKE16-NEXT:    s_branch .LBB2_12
-; GFX11-FAKE16-NEXT:  .LBB2_3:
-; GFX11-FAKE16-NEXT:    s_mov_b32 s2, 0
-; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB2_12
-; GFX11-FAKE16-NEXT:  .LBB2_4: ; %bb16
-; GFX11-FAKE16-NEXT:    s_load_b32 s0, s[16:17], 0x54
-; GFX11-FAKE16-NEXT:    s_bitcmp1_b32 s23, 0
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s9, -1, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s23, 1
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_bitcmp1_b32 s0, 0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, -1
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s8, -1, 0
-; GFX11-FAKE16-NEXT:    s_cmp_eq_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB2_8
-; GFX11-FAKE16-NEXT:  ; %bb.5: ; %bb18.preheader
-; GFX11-FAKE16-NEXT:    s_load_b128 s[28:31], s[16:17], 0x44
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_mul_hi_u32 s0, s29, s28
-; GFX11-FAKE16-NEXT:    s_mul_i32 s1, s29, s28
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_alignbit_b32 v0, s0, s1, 1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s1, 0
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s0, 1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s0, s0, s30
-; GFX11-FAKE16-NEXT:    s_mul_i32 s0, s0, s22
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_mul_i32 s0, s0, s20
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s19, s0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_lshl_b64 s[20:21], s[0:1], 1
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, s1
-; GFX11-FAKE16-NEXT:    global_load_u16 v1, v0, s[20:21]
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s24
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v1
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e64 v1, 0, 1, vcc_lo
-; GFX11-FAKE16-NEXT:    s_mov_b32 vcc_lo, 0
-; GFX11-FAKE16-NEXT:    .p2align 6
-; GFX11-FAKE16-NEXT:  .LBB2_6: ; %bb18
-; GFX11-FAKE16-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, 0xffff, s1
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v0
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s1, -1, 0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s8, s1
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, exec_lo
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s19, v2
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s1, s19, s13
-; GFX11-FAKE16-NEXT:    s_and_b32 s13, 0xffff, s0
-; GFX11-FAKE16-NEXT:    s_and_b32 s1, s1, 1
-; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s13, 0
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s13, -1, 0
-; GFX11-FAKE16-NEXT:    s_and_b32 s20, s9, exec_lo
-; GFX11-FAKE16-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s13
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s19, v2
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s13, s19, s13
-; GFX11-FAKE16-NEXT:    s_bitcmp1_b32 s13, 0
-; GFX11-FAKE16-NEXT:    s_cselect_b32 s13, 0x100, 0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_or_b32 s0, s13, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccz .LBB2_6
-; GFX11-FAKE16-NEXT:  ; %bb.7: ; %Flow
-; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
-; GFX11-FAKE16-NEXT:  .LBB2_8: ; %Flow12
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_and_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccz .LBB2_12
-; GFX11-FAKE16-NEXT:  ; %bb.9:
-; GFX11-FAKE16-NEXT:    s_xor_b32 s0, s8, -1
-; GFX11-FAKE16-NEXT:  .LBB2_10: ; %bb17
-; GFX11-FAKE16-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_and_b32 vcc_lo, exec_lo, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_vccz .LBB2_10
-; GFX11-FAKE16-NEXT:  ; %bb.11: ; %Flow6
-; GFX11-FAKE16-NEXT:    s_mov_b32 s18, -1
-; GFX11-FAKE16-NEXT:  .LBB2_12: ; %Flow11
-; GFX11-FAKE16-NEXT:    s_and_b32 s20, s2, exec_lo
-; GFX11-FAKE16-NEXT:    s_or_not1_b32 s0, s18, exec_lo
-; GFX11-FAKE16-NEXT:  .LBB2_13: ; %Flow9
-; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s3
-; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s3, s0
-; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB2_15
-; GFX11-FAKE16-NEXT:  ; %bb.14: ; %bb43
-; GFX11-FAKE16-NEXT:    s_add_u32 s8, s16, 0x58
-; GFX11-FAKE16-NEXT:    s_addc_u32 s9, s17, 0
-; GFX11-FAKE16-NEXT:    s_getpc_b64 s[0:1]
-; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, f0 at gotpcrel32@lo+4
-; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, f0 at gotpcrel32@hi+12
-; GFX11-FAKE16-NEXT:    s_mov_b32 s13, s14
-; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, s15
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_swappc_b64 s[30:31], s[0:1]
-; GFX11-FAKE16-NEXT:    s_or_b32 s20, s20, exec_lo
-; GFX11-FAKE16-NEXT:  .LBB2_15: ; %Flow14
-; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s3
-; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, s20
-; GFX11-FAKE16-NEXT:  ; %bb.16: ; %UnifiedUnreachableBlock
-; GFX11-FAKE16-NEXT:    ; divergent unreachable
-; GFX11-FAKE16-NEXT:  ; %bb.17: ; %UnifiedReturnBlock
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: f2:
+; GFX11:       ; %bb.0: ; %bb
+; GFX11-NEXT:    s_mov_b64 s[16:17], s[4:5]
+; GFX11-NEXT:    v_mov_b32_e32 v31, v0
+; GFX11-NEXT:    s_load_b32 s19, s[16:17], 0x24
+; GFX11-NEXT:    s_mov_b32 s12, s13
+; GFX11-NEXT:    s_mov_b64 s[10:11], s[6:7]
+; GFX11-NEXT:    s_mov_b64 s[6:7], s[2:3]
+; GFX11-NEXT:    v_and_b32_e32 v0, 0x3ff, v31
+; GFX11-NEXT:    s_mov_b64 s[4:5], s[0:1]
+; GFX11-NEXT:    s_mov_b32 s20, 0
+; GFX11-NEXT:    s_mov_b32 s0, -1
+; GFX11-NEXT:    s_mov_b32 s3, exec_lo
+; GFX11-NEXT:    s_mov_b32 s32, 0
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    v_mul_lo_u32 v0, s19, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_cmpx_eq_u32_e32 0, v0
+; GFX11-NEXT:    s_cbranch_execz .LBB2_13
+; GFX11-NEXT:  ; %bb.1: ; %bb14
+; GFX11-NEXT:    s_load_b128 s[20:23], s[16:17], 0x2c
+; GFX11-NEXT:    s_mov_b32 s18, 0
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_bitcmp1_b32 s21, 0
+; GFX11-NEXT:    s_cselect_b32 s24, -1, 0
+; GFX11-NEXT:    s_bitcmp0_b32 s21, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB2_3
+; GFX11-NEXT:  ; %bb.2: ; %bb15
+; GFX11-NEXT:    s_add_u32 s8, s16, 0x58
+; GFX11-NEXT:    s_addc_u32 s9, s17, 0
+; GFX11-NEXT:    s_getpc_b64 s[0:1]
+; GFX11-NEXT:    s_add_u32 s0, s0, f0 at gotpcrel32@lo+4
+; GFX11-NEXT:    s_addc_u32 s1, s1, f0 at gotpcrel32@hi+12
+; GFX11-NEXT:    s_mov_b32 s13, s14
+; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s21, s14
+; GFX11-NEXT:    s_mov_b32 s14, s15
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX11-NEXT:    s_mov_b32 s14, s21
+; GFX11-NEXT:    s_mov_b32 s2, -1
+; GFX11-NEXT:    s_cbranch_execz .LBB2_4
+; GFX11-NEXT:    s_branch .LBB2_12
+; GFX11-NEXT:  .LBB2_3:
+; GFX11-NEXT:    s_mov_b32 s2, 0
+; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccnz .LBB2_12
+; GFX11-NEXT:  .LBB2_4: ; %bb16
+; GFX11-NEXT:    s_load_b32 s0, s[16:17], 0x54
+; GFX11-NEXT:    s_bitcmp1_b32 s23, 0
+; GFX11-NEXT:    s_cselect_b32 s9, -1, 0
+; GFX11-NEXT:    s_and_b32 s1, s23, 1
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_bitcmp1_b32 s0, 0
+; GFX11-NEXT:    s_mov_b32 s0, -1
+; GFX11-NEXT:    s_cselect_b32 s8, -1, 0
+; GFX11-NEXT:    s_cmp_eq_u32 s1, 0
+; GFX11-NEXT:    s_cbranch_scc0 .LBB2_8
+; GFX11-NEXT:  ; %bb.5: ; %bb18.preheader
+; GFX11-NEXT:    s_load_b128 s[28:31], s[16:17], 0x44
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_mul_hi_u32 s0, s29, s28
+; GFX11-NEXT:    s_mul_i32 s1, s29, s28
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_alignbit_b32 v0, s0, s1, 1
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, 0
+; GFX11-NEXT:    s_or_b32 s0, s0, 1
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_lshr_b32 s0, s0, s30
+; GFX11-NEXT:    s_mul_i32 s0, s0, s22
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_mul_i32 s0, s0, s20
+; GFX11-NEXT:    s_or_b32 s0, s19, s0
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_lshl_b64 s[20:21], s[0:1], 1
+; GFX11-NEXT:    s_mov_b32 s0, s1
+; GFX11-NEXT:    global_load_u16 v1, v0, s[20:21]
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s24
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, 0, 1, vcc_lo
+; GFX11-NEXT:    s_mov_b32 vcc_lo, 0
+; GFX11-NEXT:    .p2align 6
+; GFX11-NEXT:  .LBB2_6: ; %bb18
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_and_b32 s1, 0xffff, s1
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v0
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    s_cselect_b32 s1, -1, 0
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s1
+; GFX11-NEXT:    s_and_b32 s1, s8, s1
+; GFX11-NEXT:    s_and_b32 s1, s1, exec_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s19, v2
+; GFX11-NEXT:    s_cselect_b32 s1, s19, s13
+; GFX11-NEXT:    s_and_b32 s13, 0xffff, s0
+; GFX11-NEXT:    s_and_b32 s1, s1, 1
+; GFX11-NEXT:    s_cmp_lg_u32 s13, 0
+; GFX11-NEXT:    s_cselect_b32 s13, -1, 0
+; GFX11-NEXT:    s_and_b32 s20, s9, exec_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s13
+; GFX11-NEXT:    v_readfirstlane_b32 s13, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s19, v2
+; GFX11-NEXT:    s_cselect_b32 s13, s19, s13
+; GFX11-NEXT:    s_bitcmp1_b32 s13, 0
+; GFX11-NEXT:    s_cselect_b32 s13, 0x100, 0
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_or_b32 s0, s13, s0
+; GFX11-NEXT:    s_cbranch_vccz .LBB2_6
+; GFX11-NEXT:  ; %bb.7: ; %Flow
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:  .LBB2_8: ; %Flow12
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccz .LBB2_12
+; GFX11-NEXT:  ; %bb.9:
+; GFX11-NEXT:    s_xor_b32 s0, s8, -1
+; GFX11-NEXT:  .LBB2_10: ; %bb17
+; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_and_b32 vcc_lo, exec_lo, s0
+; GFX11-NEXT:    s_cbranch_vccz .LBB2_10
+; GFX11-NEXT:  ; %bb.11: ; %Flow6
+; GFX11-NEXT:    s_mov_b32 s18, -1
+; GFX11-NEXT:  .LBB2_12: ; %Flow11
+; GFX11-NEXT:    s_and_b32 s20, s2, exec_lo
+; GFX11-NEXT:    s_or_not1_b32 s0, s18, exec_lo
+; GFX11-NEXT:  .LBB2_13: ; %Flow9
+; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s3
+; GFX11-NEXT:    s_and_saveexec_b32 s3, s0
+; GFX11-NEXT:    s_cbranch_execz .LBB2_15
+; GFX11-NEXT:  ; %bb.14: ; %bb43
+; GFX11-NEXT:    s_add_u32 s8, s16, 0x58
+; GFX11-NEXT:    s_addc_u32 s9, s17, 0
+; GFX11-NEXT:    s_getpc_b64 s[0:1]
+; GFX11-NEXT:    s_add_u32 s0, s0, f0 at gotpcrel32@lo+4
+; GFX11-NEXT:    s_addc_u32 s1, s1, f0 at gotpcrel32@hi+12
+; GFX11-NEXT:    s_mov_b32 s13, s14
+; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
+; GFX11-NEXT:    s_mov_b32 s14, s15
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_swappc_b64 s[30:31], s[0:1]
+; GFX11-NEXT:    s_or_b32 s20, s20, exec_lo
+; GFX11-NEXT:  .LBB2_15: ; %Flow14
+; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s3
+; GFX11-NEXT:    s_and_saveexec_b32 s0, s20
+; GFX11-NEXT:  ; %bb.16: ; %UnifiedUnreachableBlock
+; GFX11-NEXT:    ; divergent unreachable
+; GFX11-NEXT:  ; %bb.17: ; %UnifiedReturnBlock
+; GFX11-NEXT:    s_endpgm
 bb:
   %i = tail call i32 @llvm.amdgcn.workitem.id.x()
   %i12 = mul i32 %arg, %i
@@ -410,4 +257,3 @@ bb43:
 }
 
 attributes #0 = { noinline optnone }
-
diff --git a/llvm/test/CodeGen/AMDGPU/insert_vector_elt.v2i16.ll b/llvm/test/CodeGen/AMDGPU/insert_vector_elt.v2i16.ll
index a0ad6328b0c01..15efb2a4a3b61 100644
--- a/llvm/test/CodeGen/AMDGPU/insert_vector_elt.v2i16.ll
+++ b/llvm/test/CodeGen/AMDGPU/insert_vector_elt.v2i16.ll
@@ -105,34 +105,19 @@ define amdgpu_kernel void @s_insertelement_v2i16_0_reg(ptr addrspace(1) %out, pt
 ; CI-NEXT:    flat_store_dword v[0:1], v2
 ; CI-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: s_insertelement_v2i16_0_reg:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-TRUE16-NEXT:    s_load_b32 s4, s[4:5], 0x30
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[2:3], 0x0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s4
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_pack_lh_b32_b16 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-TRUE16-NEXT:    global_store_b32 v0, v1, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: s_insertelement_v2i16_0_reg:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-FAKE16-NEXT:    s_load_b32 s4, s[4:5], 0x30
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[2:3], 0x0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_pack_lh_b32_b16 s2, s4, s2
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-FAKE16-NEXT:    global_store_b32 v0, v1, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_insertelement_v2i16_0_reg:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
+; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x30
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_load_b32 s2, s[2:3], 0x0
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_pack_lh_b32_b16 s2, s4, s2
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
+; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-NEXT:    s_endpgm
   %vec = load <2 x i16>, ptr addrspace(4) %vec.ptr
   %vecins = insertelement <2 x i16> %vec, i16 %elt, i32 0
   store <2 x i16> %vecins, ptr addrspace(1) %out
@@ -203,42 +188,23 @@ define amdgpu_kernel void @s_insertelement_v2i16_0_multi_use_hi_reg(ptr addrspac
 ; CI-NEXT:    ;;#ASMEND
 ; CI-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: s_insertelement_v2i16_0_multi_use_hi_reg:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-TRUE16-NEXT:    s_load_b32 s4, s[4:5], 0x30
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[2:3], 0x0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s4
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s3
-; GFX11-TRUE16-NEXT:    global_store_b32 v0, v1, s[0:1]
-; GFX11-TRUE16-NEXT:    ;;#ASMSTART
-; GFX11-TRUE16-NEXT:    ; use s2
-; GFX11-TRUE16-NEXT:    ;;#ASMEND
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: s_insertelement_v2i16_0_multi_use_hi_reg:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-FAKE16-NEXT:    s_load_b32 s4, s[4:5], 0x30
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[2:3], 0x0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s2, 16
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s4, s2
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s3
-; GFX11-FAKE16-NEXT:    global_store_b32 v0, v1, s[0:1]
-; GFX11-FAKE16-NEXT:    ;;#ASMSTART
-; GFX11-FAKE16-NEXT:    ; use s2
-; GFX11-FAKE16-NEXT:    ;;#ASMEND
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_insertelement_v2i16_0_multi_use_hi_reg:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
+; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x30
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_load_b32 s2, s[2:3], 0x0
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_lshr_b32 s2, s2, 16
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s4, s2
+; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s3
+; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-NEXT:    ;;#ASMSTART
+; GFX11-NEXT:    ; use s2
+; GFX11-NEXT:    ;;#ASMEND
+; GFX11-NEXT:    s_endpgm
   %vec = load <2 x i16>, ptr addrspace(4) %vec.ptr
   %elt1 = extractelement <2 x i16> %vec, i32 1
   %vecins = insertelement <2 x i16> %vec, i16 %elt, i32 0
@@ -614,34 +580,19 @@ define amdgpu_kernel void @s_insertelement_v2i16_1_reg(ptr addrspace(1) %out, pt
 ; CI-NEXT:    flat_store_dword v[0:1], v2
 ; CI-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: s_insertelement_v2i16_1_reg:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_clause 0x1
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-TRUE16-NEXT:    s_load_b32 s4, s[4:5], 0x30
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[2:3], 0x0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s4
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s3
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-TRUE16-NEXT:    global_store_b32 v0, v1, s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: s_insertelement_v2i16_1_reg:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_clause 0x1
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-FAKE16-NEXT:    s_load_b32 s4, s[4:5], 0x30
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[2:3], 0x0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s4
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
-; GFX11-FAKE16-NEXT:    global_store_b32 v0, v1, s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: s_insertelement_v2i16_1_reg:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_clause 0x1
+; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
+; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x30
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_load_b32 s2, s[2:3], 0x0
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s2, s4
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
+; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-NEXT:    s_endpgm
   %vec = load <2 x i16>, ptr addrspace(4) %vec.ptr
   %vecins = insertelement <2 x i16> %vec, i16 %elt, i32 1
   store <2 x i16> %vecins, ptr addrspace(1) %out
@@ -2344,49 +2295,26 @@ define amdgpu_kernel void @v_insertelement_v4i16_dynamic_vgpr(ptr addrspace(1) %
 ; CI-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; CI-NEXT:    s_endpgm
 ;
-; GFX11-TRUE16-LABEL: v_insertelement_v4i16_dynamic_vgpr:
-; GFX11-TRUE16:       ; %bb.0:
-; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-TRUE16-NEXT:    global_load_b32 v2, v[0:1], off glc dlc
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-TRUE16-NEXT:    s_load_b32 s4, s[4:5], 0x10
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
-; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    global_load_b64 v[0:1], v4, s[2:3]
-; GFX11-TRUE16-NEXT:    s_mov_b32 s2, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s2
-; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 4, v2
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshlrev_b64 v[2:3], v2, 0xffff
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v1, v3, s2, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_bfi_b32 v0, v2, s2, v0
-; GFX11-TRUE16-NEXT:    global_store_b64 v4, v[0:1], s[0:1]
-; GFX11-TRUE16-NEXT:    s_endpgm
-;
-; GFX11-FAKE16-LABEL: v_insertelement_v4i16_dynamic_vgpr:
-; GFX11-FAKE16:       ; %bb.0:
-; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
-; GFX11-FAKE16-NEXT:    global_load_b32 v2, v[0:1], off glc dlc
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
-; GFX11-FAKE16-NEXT:    s_load_b32 s4, s[4:5], 0x10
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
-; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    global_load_b64 v[0:1], v4, s[2:3]
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s4, s4
-; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 4, v2
-; GFX11-FAKE16-NEXT:    v_lshlrev_b64 v[2:3], v2, 0xffff
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v1, v3, s2, v1
-; GFX11-FAKE16-NEXT:    v_bfi_b32 v0, v2, s2, v0
-; GFX11-FAKE16-NEXT:    global_store_b64 v4, v[0:1], s[0:1]
-; GFX11-FAKE16-NEXT:    s_endpgm
+; GFX11-LABEL: v_insertelement_v4i16_dynamic_vgpr:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
+; GFX11-NEXT:    global_load_b32 v2, v[0:1], off glc dlc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
+; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x10
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
+; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    global_load_b64 v[0:1], v4, s[2:3]
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s4, s4
+; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 4, v2
+; GFX11-NEXT:    v_lshlrev_b64 v[2:3], v2, 0xffff
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_bfi_b32 v1, v3, s2, v1
+; GFX11-NEXT:    v_bfi_b32 v0, v2, s2, v0
+; GFX11-NEXT:    global_store_b64 v4, v[0:1], s[0:1]
+; GFX11-NEXT:    s_endpgm
   %tid = call i32 @llvm.amdgcn.workitem.id.x() #1
   %tid.ext = sext i32 %tid to i64
   %in.gep = getelementptr inbounds <4 x i16>, ptr addrspace(1) %in, i64 %tid.ext
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.intersect_ray.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.intersect_ray.ll
index 5bbb0e5563cc2..2c1b68278037a 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.intersect_ray.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.intersect_ray.ll
@@ -72,98 +72,47 @@ define amdgpu_ps <4 x float> @image_bvh_intersect_ray_a16(i32 inreg %node_ptr, f
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
-; GFX11-TRUE16-LABEL: image_bvh_intersect_ray_a16:
-; GFX11-TRUE16:       ; %bb.0: ; %main_body
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s2, s7, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s5, 16
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s3, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s5
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s8
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s5
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s6
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_mov_b32 v4, s2
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s3 :: v_dual_mov_b32 v6, s0
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v7, s1
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s10
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s9
-; GFX11-TRUE16-NEXT:    image_bvh_intersect_ray v[0:3], [v6, v7, v[0:2], v[3:5]], s[12:15] a16
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-FAKE16-LABEL: image_bvh_intersect_ray_a16:
-; GFX11-FAKE16:       ; %bb.0: ; %main_body
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s2, s7, 16
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s5, 16
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s0 :: v_dual_mov_b32 v7, s1
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s3, s2
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s7
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s3
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s6, s8
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s2 :: v_dual_mov_b32 v5, s4
-; GFX11-FAKE16-NEXT:    s_mov_b32 s15, s12
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, s11
-; GFX11-FAKE16-NEXT:    s_mov_b32 s13, s10
-; GFX11-FAKE16-NEXT:    s_mov_b32 s12, s9
-; GFX11-FAKE16-NEXT:    image_bvh_intersect_ray v[0:3], [v6, v7, v[0:2], v[3:5]], s[12:15] a16
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    ; return to shader part epilog
-;
-; GFX12-SDAG-TRUE16-LABEL: image_bvh_intersect_ray_a16:
-; GFX12-SDAG-TRUE16:       ; %bb.0: ; %main_body
-; GFX12-SDAG-TRUE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
-; GFX12-SDAG-TRUE16-NEXT:    s_lshr_b32 s2, s7, 16
-; GFX12-SDAG-TRUE16-NEXT:    s_lshr_b32 s3, s5, 16
-; GFX12-SDAG-TRUE16-NEXT:    v_mov_b32_e32 v2, s4
-; GFX12-SDAG-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s3, s2
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s3, s6
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s4, s5
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s5, s7
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s6, s8
-; GFX12-SDAG-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s5
-; GFX12-SDAG-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s6
-; GFX12-SDAG-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-TRUE16-NEXT:    v_dual_mov_b32 v3, s4 :: v_dual_mov_b32 v4, s2
-; GFX12-SDAG-TRUE16-NEXT:    v_dual_mov_b32 v5, s3 :: v_dual_mov_b32 v6, s0
-; GFX12-SDAG-TRUE16-NEXT:    v_mov_b32_e32 v7, s1
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s15, s12
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s14, s11
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s13, s10
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s12, s9
-; GFX12-SDAG-TRUE16-NEXT:    image_bvh_intersect_ray v[0:3], [v6, v7, v[0:2], v[3:5]], s[12:15] a16
-; GFX12-SDAG-TRUE16-NEXT:    s_wait_bvhcnt 0x0
-; GFX12-SDAG-TRUE16-NEXT:    ; return to shader part epilog
+; GFX11-LABEL: image_bvh_intersect_ray_a16:
+; GFX11:       ; %bb.0: ; %main_body
+; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-NEXT:    s_lshr_b32 s2, s7, 16
+; GFX11-NEXT:    s_lshr_b32 s3, s5, 16
+; GFX11-NEXT:    v_dual_mov_b32 v6, s0 :: v_dual_mov_b32 v7, s1
+; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s3, s2
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s5, s7
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s3
+; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s6, s8
+; GFX11-NEXT:    v_dual_mov_b32 v4, s2 :: v_dual_mov_b32 v5, s4
+; GFX11-NEXT:    s_mov_b32 s15, s12
+; GFX11-NEXT:    s_mov_b32 s14, s11
+; GFX11-NEXT:    s_mov_b32 s13, s10
+; GFX11-NEXT:    s_mov_b32 s12, s9
+; GFX11-NEXT:    image_bvh_intersect_ray v[0:3], [v6, v7, v[0:2], v[3:5]], s[12:15] a16
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    ; return to shader part epilog
 ;
-; GFX12-SDAG-FAKE16-LABEL: image_bvh_intersect_ray_a16:
-; GFX12-SDAG-FAKE16:       ; %bb.0: ; %main_body
-; GFX12-SDAG-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
-; GFX12-SDAG-FAKE16-NEXT:    s_lshr_b32 s2, s7, 16
-; GFX12-SDAG-FAKE16-NEXT:    s_lshr_b32 s3, s5, 16
-; GFX12-SDAG-FAKE16-NEXT:    v_dual_mov_b32 v6, s0 :: v_dual_mov_b32 v7, s1
-; GFX12-SDAG-FAKE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s3, s2
-; GFX12-SDAG-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s5, s7
-; GFX12-SDAG-FAKE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-FAKE16-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s3
-; GFX12-SDAG-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s6, s8
-; GFX12-SDAG-FAKE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-FAKE16-NEXT:    v_dual_mov_b32 v4, s2 :: v_dual_mov_b32 v5, s4
-; GFX12-SDAG-FAKE16-NEXT:    s_mov_b32 s15, s12
-; GFX12-SDAG-FAKE16-NEXT:    s_mov_b32 s14, s11
-; GFX12-SDAG-FAKE16-NEXT:    s_mov_b32 s13, s10
-; GFX12-SDAG-FAKE16-NEXT:    s_mov_b32 s12, s9
-; GFX12-SDAG-FAKE16-NEXT:    image_bvh_intersect_ray v[0:3], [v6, v7, v[0:2], v[3:5]], s[12:15] a16
-; GFX12-SDAG-FAKE16-NEXT:    s_wait_bvhcnt 0x0
-; GFX12-SDAG-FAKE16-NEXT:    ; return to shader part epilog
+; GFX12-SDAG-LABEL: image_bvh_intersect_ray_a16:
+; GFX12-SDAG:       ; %bb.0: ; %main_body
+; GFX12-SDAG-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX12-SDAG-NEXT:    s_lshr_b32 s2, s7, 16
+; GFX12-SDAG-NEXT:    s_lshr_b32 s3, s5, 16
+; GFX12-SDAG-NEXT:    v_dual_mov_b32 v6, s0 :: v_dual_mov_b32 v7, s1
+; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
+; GFX12-SDAG-NEXT:    s_pack_ll_b32_b16 s2, s3, s2
+; GFX12-SDAG-NEXT:    s_pack_ll_b32_b16 s3, s5, s7
+; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
+; GFX12-SDAG-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s3
+; GFX12-SDAG-NEXT:    s_pack_ll_b32_b16 s4, s6, s8
+; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
+; GFX12-SDAG-NEXT:    v_dual_mov_b32 v4, s2 :: v_dual_mov_b32 v5, s4
+; GFX12-SDAG-NEXT:    s_mov_b32 s15, s12
+; GFX12-SDAG-NEXT:    s_mov_b32 s14, s11
+; GFX12-SDAG-NEXT:    s_mov_b32 s13, s10
+; GFX12-SDAG-NEXT:    s_mov_b32 s12, s9
+; GFX12-SDAG-NEXT:    image_bvh_intersect_ray v[0:3], [v6, v7, v[0:2], v[3:5]], s[12:15] a16
+; GFX12-SDAG-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-SDAG-NEXT:    ; return to shader part epilog
 ;
 ; GFX12-GISEL-TRUE16-LABEL: image_bvh_intersect_ray_a16:
 ; GFX12-GISEL-TRUE16:       ; %bb.0: ; %main_body
@@ -263,100 +212,49 @@ define amdgpu_ps <4 x float> @image_bvh64_intersect_ray_a16(i64 inreg %node_ptr,
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    ; return to shader part epilog
 ;
-; GFX11-TRUE16-LABEL: image_bvh64_intersect_ray_a16:
-; GFX11-TRUE16:       ; %bb.0: ; %main_body
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s3 :: v_dual_mov_b32 v1, s4
-; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v6, s0
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s0, s8, 16
-; GFX11-TRUE16-NEXT:    s_lshr_b32 s3, s6, 16
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v7, s1
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s3, s0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s7
-; GFX11-TRUE16-NEXT:    s_mov_b32 s3, s6
-; GFX11-TRUE16-NEXT:    s_mov_b32 s4, s8
-; GFX11-TRUE16-NEXT:    s_mov_b32 s5, s9
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v4, s0
-; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s1 :: v_dual_mov_b32 v8, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s13
-; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s12
-; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s11
-; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s10
-; GFX11-TRUE16-NEXT:    image_bvh64_intersect_ray v[0:3], [v[6:7], v8, v[0:2], v[3:5]], s[12:15] a16
-; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-FAKE16-LABEL: image_bvh64_intersect_ray_a16:
-; GFX11-FAKE16:       ; %bb.0: ; %main_body
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s3 :: v_dual_mov_b32 v1, s4
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v7, s1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s3, s6, 16
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s6, s8
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s0 :: v_dual_mov_b32 v3, s1
-; GFX11-FAKE16-NEXT:    s_lshr_b32 s0, s8, 16
-; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, s2
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s3, s0
-; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s7, s9
-; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s3
-; GFX11-FAKE16-NEXT:    s_mov_b32 s15, s13
-; GFX11-FAKE16-NEXT:    s_mov_b32 s14, s12
-; GFX11-FAKE16-NEXT:    s_mov_b32 s13, s11
-; GFX11-FAKE16-NEXT:    s_mov_b32 s12, s10
-; GFX11-FAKE16-NEXT:    image_bvh64_intersect_ray v[0:3], [v[6:7], v8, v[0:2], v[3:5]], s[12:15] a16
-; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FAKE16-NEXT:    ; return to shader part epilog
-;
-; GFX12-SDAG-TRUE16-LABEL: image_bvh64_intersect_ray_a16:
-; GFX12-SDAG-TRUE16:       ; %bb.0: ; %main_body
-; GFX12-SDAG-TRUE16-NEXT:    v_dual_mov_b32 v0, s3 :: v_dual_mov_b32 v1, s4
-; GFX12-SDAG-TRUE16-NEXT:    v_mov_b32_e32 v6, s0
-; GFX12-SDAG-TRUE16-NEXT:    s_lshr_b32 s0, s8, 16
-; GFX12-SDAG-TRUE16-NEXT:    s_lshr_b32 s3, s6, 16
-; GFX12-SDAG-TRUE16-NEXT:    v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v7, s1
-; GFX12-SDAG-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-TRUE16-NEXT:    s_pack_ll_b32_b16 s0, s3, s0
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s1, s7
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s3, s6
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s4, s8
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s5, s9
-; GFX12-SDAG-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX12-SDAG-TRUE16-NEXT:    s_pack_ll_b32_b16 s1, s1, s5
-; GFX12-SDAG-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-TRUE16-NEXT:    v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v4, s0
-; GFX12-SDAG-TRUE16-NEXT:    v_dual_mov_b32 v5, s1 :: v_dual_mov_b32 v8, s2
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s15, s13
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s14, s12
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s13, s11
-; GFX12-SDAG-TRUE16-NEXT:    s_mov_b32 s12, s10
-; GFX12-SDAG-TRUE16-NEXT:    image_bvh64_intersect_ray v[0:3], [v[6:7], v8, v[0:2], v[3:5]], s[12:15] a16
-; GFX12-SDAG-TRUE16-NEXT:    s_wait_bvhcnt 0x0
-; GFX12-SDAG-TRUE16-NEXT:    ; return to shader part epilog
+; GFX11-LABEL: image_bvh64_intersect_ray_a16:
+; GFX11:       ; %bb.0: ; %main_body
+; GFX11-NEXT:    v_dual_mov_b32 v0, s3 :: v_dual_mov_b32 v1, s4
+; GFX11-NEXT:    v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v7, s1
+; GFX11-NEXT:    s_lshr_b32 s3, s6, 16
+; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s6, s8
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v6, s0 :: v_dual_mov_b32 v3, s1
+; GFX11-NEXT:    s_lshr_b32 s0, s8, 16
+; GFX11-NEXT:    v_mov_b32_e32 v8, s2
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s3, s0
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s7, s9
+; GFX11-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s3
+; GFX11-NEXT:    s_mov_b32 s15, s13
+; GFX11-NEXT:    s_mov_b32 s14, s12
+; GFX11-NEXT:    s_mov_b32 s13, s11
+; GFX11-NEXT:    s_mov_b32 s12, s10
+; GFX11-NEXT:    image_bvh64_intersect_ray v[0:3], [v[6:7], v8, v[0:2], v[3:5]], s[12:15] a16
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    ; return to shader part epilog
 ;
-; GFX12-SDAG-FAKE16-LABEL: image_bvh64_intersect_ray_a16:
-; GFX12-SDAG-FAKE16:       ; %bb.0: ; %main_body
-; GFX12-SDAG-FAKE16-NEXT:    v_dual_mov_b32 v0, s3 :: v_dual_mov_b32 v1, s4
-; GFX12-SDAG-FAKE16-NEXT:    v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v7, s1
-; GFX12-SDAG-FAKE16-NEXT:    s_lshr_b32 s3, s6, 16
-; GFX12-SDAG-FAKE16-NEXT:    s_pack_ll_b32_b16 s1, s6, s8
-; GFX12-SDAG-FAKE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-FAKE16-NEXT:    v_dual_mov_b32 v6, s0 :: v_dual_mov_b32 v3, s1
-; GFX12-SDAG-FAKE16-NEXT:    s_lshr_b32 s0, s8, 16
-; GFX12-SDAG-FAKE16-NEXT:    v_mov_b32_e32 v8, s2
-; GFX12-SDAG-FAKE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-FAKE16-NEXT:    s_pack_ll_b32_b16 s0, s3, s0
-; GFX12-SDAG-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s7, s9
-; GFX12-SDAG-FAKE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-SDAG-FAKE16-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s3
-; GFX12-SDAG-FAKE16-NEXT:    s_mov_b32 s15, s13
-; GFX12-SDAG-FAKE16-NEXT:    s_mov_b32 s14, s12
-; GFX12-SDAG-FAKE16-NEXT:    s_mov_b32 s13, s11
-; GFX12-SDAG-FAKE16-NEXT:    s_mov_b32 s12, s10
-; GFX12-SDAG-FAKE16-NEXT:    image_bvh64_intersect_ray v[0:3], [v[6:7], v8, v[0:2], v[3:5]], s[12:15] a16
-; GFX12-SDAG-FAKE16-NEXT:    s_wait_bvhcnt 0x0
-; GFX12-SDAG-FAKE16-NEXT:    ; return to shader part epilog
+; GFX12-SDAG-LABEL: image_bvh64_intersect_ray_a16:
+; GFX12-SDAG:       ; %bb.0: ; %main_body
+; GFX12-SDAG-NEXT:    v_dual_mov_b32 v0, s3 :: v_dual_mov_b32 v1, s4
+; GFX12-SDAG-NEXT:    v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v7, s1
+; GFX12-SDAG-NEXT:    s_lshr_b32 s3, s6, 16
+; GFX12-SDAG-NEXT:    s_pack_ll_b32_b16 s1, s6, s8
+; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
+; GFX12-SDAG-NEXT:    v_dual_mov_b32 v6, s0 :: v_dual_mov_b32 v3, s1
+; GFX12-SDAG-NEXT:    s_lshr_b32 s0, s8, 16
+; GFX12-SDAG-NEXT:    v_mov_b32_e32 v8, s2
+; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
+; GFX12-SDAG-NEXT:    s_pack_ll_b32_b16 s0, s3, s0
+; GFX12-SDAG-NEXT:    s_pack_ll_b32_b16 s3, s7, s9
+; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
+; GFX12-SDAG-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s3
+; GFX12-SDAG-NEXT:    s_mov_b32 s15, s13
+; GFX12-SDAG-NEXT:    s_mov_b32 s14, s12
+; GFX12-SDAG-NEXT:    s_mov_b32 s13, s11
+; GFX12-SDAG-NEXT:    s_mov_b32 s12, s10
+; GFX12-SDAG-NEXT:    image_bvh64_intersect_ray v[0:3], [v[6:7], v8, v[0:2], v[3:5]], s[12:15] a16
+; GFX12-SDAG-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-SDAG-NEXT:    ; return to shader part epilog
 ;
 ; GFX12-GISEL-TRUE16-LABEL: image_bvh64_intersect_ray_a16:
 ; GFX12-GISEL-TRUE16:       ; %bb.0: ; %main_body
@@ -1031,3 +929,8 @@ main_body:
 }
 
 declare i32 @llvm.amdgcn.workitem.id.x()
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; GFX11-FAKE16: {{.*}}
+; GFX11-TRUE16: {{.*}}
+; GFX12-SDAG-FAKE16: {{.*}}
+; GFX12-SDAG-TRUE16: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.ptr.tbuffer.load.d16.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.ptr.tbuffer.load.d16.ll
index 3fd100d4a7c89..381924ebbd0ed 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.ptr.tbuffer.load.d16.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.ptr.tbuffer.load.d16.ll
@@ -3,8 +3,8 @@
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx810 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX10-PACKED %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED,GFX11-PACKED-TRUE16 %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED,GFX11-PACKED-FAKE16 %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s
 
 define amdgpu_ps half @tbuffer_load_d16_x(ptr addrspace(8) inreg %rsrc) {
 ; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_x:
@@ -91,19 +91,12 @@ define amdgpu_ps half @tbuffer_load_d16_xyz(ptr addrspace(8) inreg %rsrc) {
 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, v1
 ; GFX10-PACKED-NEXT:    ; return to shader part epilog
 ;
-; GFX11-PACKED-TRUE16-LABEL: tbuffer_load_d16_xyz:
-; GFX11-PACKED-TRUE16:       ; %bb.0: ; %main_body
-; GFX11-PACKED-TRUE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]
-; GFX11-PACKED-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-PACKED-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.l
-; GFX11-PACKED-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-PACKED-FAKE16-LABEL: tbuffer_load_d16_xyz:
-; GFX11-PACKED-FAKE16:       ; %bb.0: ; %main_body
-; GFX11-PACKED-FAKE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]
-; GFX11-PACKED-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-PACKED-FAKE16-NEXT:    v_mov_b32_e32 v0, v1
-; GFX11-PACKED-FAKE16-NEXT:    ; return to shader part epilog
+; GFX11-PACKED-LABEL: tbuffer_load_d16_xyz:
+; GFX11-PACKED:       ; %bb.0: ; %main_body
+; GFX11-PACKED-NEXT:    tbuffer_load_d16_format_xyz v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]
+; GFX11-PACKED-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, v1
+; GFX11-PACKED-NEXT:    ; return to shader part epilog
 main_body:
   %data = call <3 x half> @llvm.amdgcn.raw.ptr.tbuffer.load.v3f16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 22, i32 0)
   %elt = extractelement <3 x half> %data, i32 2
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.tbuffer.load.d16.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.tbuffer.load.d16.ll
index e08123608176c..bbac914714aba 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.tbuffer.load.d16.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.tbuffer.load.d16.ll
@@ -3,10 +3,10 @@
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx810 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX10-PACKED %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED,GFX11-PACKED-TRUE16 %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED,GFX11-PACKED-FAKE16 %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX12-PACKED,GFX12-PACKED-TRUE16 %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX12-PACKED,GFX12-PACKED-FAKE16 %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX12-PACKED %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 -amdgpu-enable-delay-alu=0 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX12-PACKED %s
 
 define amdgpu_ps half @tbuffer_load_d16_x(<4 x i32> inreg %rsrc) {
 ; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_x:
@@ -106,33 +106,19 @@ define amdgpu_ps half @tbuffer_load_d16_xyz(<4 x i32> inreg %rsrc) {
 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, v1
 ; GFX10-PACKED-NEXT:    ; return to shader part epilog
 ;
-; GFX11-PACKED-TRUE16-LABEL: tbuffer_load_d16_xyz:
-; GFX11-PACKED-TRUE16:       ; %bb.0: ; %main_body
-; GFX11-PACKED-TRUE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]
-; GFX11-PACKED-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-PACKED-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.l
-; GFX11-PACKED-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-PACKED-FAKE16-LABEL: tbuffer_load_d16_xyz:
-; GFX11-PACKED-FAKE16:       ; %bb.0: ; %main_body
-; GFX11-PACKED-FAKE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]
-; GFX11-PACKED-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-PACKED-FAKE16-NEXT:    v_mov_b32_e32 v0, v1
-; GFX11-PACKED-FAKE16-NEXT:    ; return to shader part epilog
-;
-; GFX12-PACKED-TRUE16-LABEL: tbuffer_load_d16_xyz:
-; GFX12-PACKED-TRUE16:       ; %bb.0: ; %main_body
-; GFX12-PACKED-TRUE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], off, s[0:3], null format:[BUF_FMT_32_FLOAT]
-; GFX12-PACKED-TRUE16-NEXT:    s_wait_loadcnt 0x0
-; GFX12-PACKED-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.l
-; GFX12-PACKED-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX12-PACKED-FAKE16-LABEL: tbuffer_load_d16_xyz:
-; GFX12-PACKED-FAKE16:       ; %bb.0: ; %main_body
-; GFX12-PACKED-FAKE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], off, s[0:3], null format:[BUF_FMT_32_FLOAT]
-; GFX12-PACKED-FAKE16-NEXT:    s_wait_loadcnt 0x0
-; GFX12-PACKED-FAKE16-NEXT:    v_mov_b32_e32 v0, v1
-; GFX12-PACKED-FAKE16-NEXT:    ; return to shader part epilog
+; GFX11-PACKED-LABEL: tbuffer_load_d16_xyz:
+; GFX11-PACKED:       ; %bb.0: ; %main_body
+; GFX11-PACKED-NEXT:    tbuffer_load_d16_format_xyz v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT]
+; GFX11-PACKED-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, v1
+; GFX11-PACKED-NEXT:    ; return to shader part epilog
+;
+; GFX12-PACKED-LABEL: tbuffer_load_d16_xyz:
+; GFX12-PACKED:       ; %bb.0: ; %main_body
+; GFX12-PACKED-NEXT:    tbuffer_load_d16_format_xyz v[0:1], off, s[0:3], null format:[BUF_FMT_32_FLOAT]
+; GFX12-PACKED-NEXT:    s_wait_loadcnt 0x0
+; GFX12-PACKED-NEXT:    v_mov_b32_e32 v0, v1
+; GFX12-PACKED-NEXT:    ; return to shader part epilog
 main_body:
   %data = call <3 x half> @llvm.amdgcn.raw.tbuffer.load.v3f16(<4 x i32> %rsrc, i32 0, i32 0, i32 22, i32 0)
   %elt = extractelement <3 x half> %data, i32 2
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.struct.ptr.tbuffer.load.d16.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.struct.ptr.tbuffer.load.d16.ll
index 98f3a4dd1c690..2f26743e470a7 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.struct.ptr.tbuffer.load.d16.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.struct.ptr.tbuffer.load.d16.ll
@@ -3,8 +3,8 @@
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx810 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX10-PACKED %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED,GFX11-PACKED-TRUE16 %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED,GFX11-PACKED-FAKE16 %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s
 
 define amdgpu_ps half @tbuffer_load_d16_x(ptr addrspace(8) inreg %rsrc) {
 ; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_x:
@@ -102,21 +102,13 @@ define amdgpu_ps half @tbuffer_load_d16_xyz(ptr addrspace(8) inreg %rsrc) {
 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, v1
 ; GFX10-PACKED-NEXT:    ; return to shader part epilog
 ;
-; GFX11-PACKED-TRUE16-LABEL: tbuffer_load_d16_xyz:
-; GFX11-PACKED-TRUE16:       ; %bb.0: ; %main_body
-; GFX11-PACKED-TRUE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-PACKED-TRUE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
-; GFX11-PACKED-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-PACKED-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.l
-; GFX11-PACKED-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-PACKED-FAKE16-LABEL: tbuffer_load_d16_xyz:
-; GFX11-PACKED-FAKE16:       ; %bb.0: ; %main_body
-; GFX11-PACKED-FAKE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-PACKED-FAKE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
-; GFX11-PACKED-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-PACKED-FAKE16-NEXT:    v_mov_b32_e32 v0, v1
-; GFX11-PACKED-FAKE16-NEXT:    ; return to shader part epilog
+; GFX11-PACKED-LABEL: tbuffer_load_d16_xyz:
+; GFX11-PACKED:       ; %bb.0: ; %main_body
+; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, 0
+; GFX11-PACKED-NEXT:    tbuffer_load_d16_format_xyz v[0:1], v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
+; GFX11-PACKED-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, v1
+; GFX11-PACKED-NEXT:    ; return to shader part epilog
 main_body:
   %data = call <3 x half> @llvm.amdgcn.struct.ptr.tbuffer.load.v3f16(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 0, i32 22, i32 0)
   %elt = extractelement <3 x half> %data, i32 2
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.struct.tbuffer.load.d16.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.struct.tbuffer.load.d16.ll
index 3328b40d81d5e..f93e1881d950b 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.struct.tbuffer.load.d16.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.struct.tbuffer.load.d16.ll
@@ -3,10 +3,10 @@
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx810 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=PREGFX10-PACKED %s
 ; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX10-PACKED %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED,GFX11-PACKED-TRUE16 %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED,GFX11-PACKED-FAKE16 %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX12-PACKED,GFX12-PACKED-TRUE16 %s
-; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX12-PACKED,GFX12-PACKED-FAKE16 %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX11-PACKED %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX12-PACKED %s
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 -verify-machineinstrs | FileCheck -enable-var-scope -check-prefixes=GFX12-PACKED %s
 
 define amdgpu_ps half @tbuffer_load_d16_x(<4 x i32> inreg %rsrc) {
 ; PREGFX10-UNPACKED-LABEL: tbuffer_load_d16_x:
@@ -119,37 +119,21 @@ define amdgpu_ps half @tbuffer_load_d16_xyz(<4 x i32> inreg %rsrc) {
 ; GFX10-PACKED-NEXT:    v_mov_b32_e32 v0, v1
 ; GFX10-PACKED-NEXT:    ; return to shader part epilog
 ;
-; GFX11-PACKED-TRUE16-LABEL: tbuffer_load_d16_xyz:
-; GFX11-PACKED-TRUE16:       ; %bb.0: ; %main_body
-; GFX11-PACKED-TRUE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-PACKED-TRUE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
-; GFX11-PACKED-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-PACKED-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.l
-; GFX11-PACKED-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX11-PACKED-FAKE16-LABEL: tbuffer_load_d16_xyz:
-; GFX11-PACKED-FAKE16:       ; %bb.0: ; %main_body
-; GFX11-PACKED-FAKE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-PACKED-FAKE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
-; GFX11-PACKED-FAKE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-PACKED-FAKE16-NEXT:    v_mov_b32_e32 v0, v1
-; GFX11-PACKED-FAKE16-NEXT:    ; return to shader part epilog
-;
-; GFX12-PACKED-TRUE16-LABEL: tbuffer_load_d16_xyz:
-; GFX12-PACKED-TRUE16:       ; %bb.0: ; %main_body
-; GFX12-PACKED-TRUE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX12-PACKED-TRUE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], v0, s[0:3], null format:[BUF_FMT_32_FLOAT] idxen
-; GFX12-PACKED-TRUE16-NEXT:    s_wait_loadcnt 0x0
-; GFX12-PACKED-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.l
-; GFX12-PACKED-TRUE16-NEXT:    ; return to shader part epilog
-;
-; GFX12-PACKED-FAKE16-LABEL: tbuffer_load_d16_xyz:
-; GFX12-PACKED-FAKE16:       ; %bb.0: ; %main_body
-; GFX12-PACKED-FAKE16-NEXT:    v_mov_b32_e32 v0, 0
-; GFX12-PACKED-FAKE16-NEXT:    tbuffer_load_d16_format_xyz v[0:1], v0, s[0:3], null format:[BUF_FMT_32_FLOAT] idxen
-; GFX12-PACKED-FAKE16-NEXT:    s_wait_loadcnt 0x0
-; GFX12-PACKED-FAKE16-NEXT:    v_mov_b32_e32 v0, v1
-; GFX12-PACKED-FAKE16-NEXT:    ; return to shader part epilog
+; GFX11-PACKED-LABEL: tbuffer_load_d16_xyz:
+; GFX11-PACKED:       ; %bb.0: ; %main_body
+; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, 0
+; GFX11-PACKED-NEXT:    tbuffer_load_d16_format_xyz v[0:1], v0, s[0:3], 0 format:[BUF_FMT_32_FLOAT] idxen
+; GFX11-PACKED-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-PACKED-NEXT:    v_mov_b32_e32 v0, v1
+; GFX11-PACKED-NEXT:    ; return to shader part epilog
+;
+; GFX12-PACKED-LABEL: tbuffer_load_d16_xyz:
+; GFX12-PACKED:       ; %bb.0: ; %main_body
+; GFX12-PACKED-NEXT:    v_mov_b32_e32 v0, 0
+; GFX12-PACKED-NEXT:    tbuffer_load_d16_format_xyz v[0:1], v0, s[0:3], null format:[BUF_FMT_32_FLOAT] idxen
+; GFX12-PACKED-NEXT:    s_wait_loadcnt 0x0
+; GFX12-PACKED-NEXT:    v_mov_b32_e32 v0, v1
+; GFX12-PACKED-NEXT:    ; return to shader part epilog
 main_body:
   %data = call <3 x half> @llvm.amdgcn.struct.tbuffer.load.v3f16(<4 x i32> %rsrc, i32 0, i32 0, i32 0, i32 22, i32 0)
   %elt = extractelement <3 x half> %data, i32 2
diff --git a/llvm/test/CodeGen/AMDGPU/load-constant-i16.ll b/llvm/test/CodeGen/AMDGPU/load-constant-i16.ll
index 8e312a0e195ff..58a41226fde45 100644
--- a/llvm/test/CodeGen/AMDGPU/load-constant-i16.ll
+++ b/llvm/test/CodeGen/AMDGPU/load-constant-i16.ll
@@ -5909,36 +5909,20 @@ define amdgpu_kernel void @constant_zextload_v2i16_to_v2i64(ptr addrspace(1) %ou
 ; EG-NEXT:     LSHR * T5.X, KC0[2].Y, literal.y,
 ; EG-NEXT:    65535(9.183409e-41), 2(2.802597e-45)
 ;
-; GFX12-TRUE16-LABEL: constant_zextload_v2i16_to_v2i64:
-; GFX12-TRUE16:       ; %bb.0:
-; GFX12-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    s_load_b32 s2, s[2:3], 0x0
-; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    s_mov_b32 s3, s2
-; GFX12-TRUE16-NEXT:    s_pack_hl_b32_b16 s2, s2, 0
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s2
-; GFX12-TRUE16-NEXT:    s_and_b32 s3, 0xffff, s3
-; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v0, s3 :: v_dual_mov_b32 v3, v1
-; GFX12-TRUE16-NEXT:    global_store_b128 v1, v[0:3], s[0:1]
-; GFX12-TRUE16-NEXT:    s_endpgm
-;
-; GFX12-FAKE16-LABEL: constant_zextload_v2i16_to_v2i64:
-; GFX12-FAKE16:       ; %bb.0:
-; GFX12-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-FAKE16-NEXT:    s_load_b32 s2, s[2:3], 0x0
-; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-FAKE16-NEXT:    s_and_b32 s3, 0xffff, s2
-; GFX12-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX12-FAKE16-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s3
-; GFX12-FAKE16-NEXT:    s_pack_hl_b32_b16 s2, s2, 0
-; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, v1
-; GFX12-FAKE16-NEXT:    global_store_b128 v1, v[0:3], s[0:1]
-; GFX12-FAKE16-NEXT:    s_endpgm
+; GFX12-LABEL: constant_zextload_v2i16_to_v2i64:
+; GFX12:       ; %bb.0:
+; GFX12-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
+; GFX12-NEXT:    s_wait_kmcnt 0x0
+; GFX12-NEXT:    s_load_b32 s2, s[2:3], 0x0
+; GFX12-NEXT:    s_wait_kmcnt 0x0
+; GFX12-NEXT:    s_and_b32 s3, 0xffff, s2
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX12-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s3
+; GFX12-NEXT:    s_pack_hl_b32_b16 s2, s2, 0
+; GFX12-NEXT:    s_wait_alu 0xfffe
+; GFX12-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, v1
+; GFX12-NEXT:    global_store_b128 v1, v[0:3], s[0:1]
+; GFX12-NEXT:    s_endpgm
   %load = load <2 x i16>, ptr addrspace(4) %in
   %ext = zext <2 x i16> %load to <2 x i64>
   store <2 x i64> %ext, ptr addrspace(1) %out
@@ -6154,47 +6138,26 @@ define amdgpu_kernel void @constant_zextload_v4i16_to_v4i64(ptr addrspace(1) %ou
 ; EG-NEXT:     LSHR * T8.X, PV.W, literal.x,
 ; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)
 ;
-; GFX12-TRUE16-LABEL: constant_zextload_v4i16_to_v4i64:
-; GFX12-TRUE16:       ; %bb.0:
-; GFX12-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    s_load_b64 s[2:3], s[2:3], 0x0
-; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    s_mov_b32 s4, s2
-; GFX12-TRUE16-NEXT:    s_pack_hl_b32_b16 s2, s2, 0
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s2
-; GFX12-TRUE16-NEXT:    s_and_b32 s4, 0xffff, s4
-; GFX12-TRUE16-NEXT:    s_pack_hl_b32_b16 s2, s3, 0
-; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v3, v1
-; GFX12-TRUE16-NEXT:    s_and_b32 s3, 0xffff, s3
-; GFX12-TRUE16-NEXT:    global_store_b128 v1, v[0:3], s[0:1]
-; GFX12-TRUE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v0, s3
-; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v2, s2
-; GFX12-TRUE16-NEXT:    global_store_b128 v1, v[0:3], s[0:1] offset:16
-; GFX12-TRUE16-NEXT:    s_endpgm
-;
-; GFX12-FAKE16-LABEL: constant_zextload_v4i16_to_v4i64:
-; GFX12-FAKE16:       ; %bb.0:
-; GFX12-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-FAKE16-NEXT:    s_load_b64 s[2:3], s[2:3], 0x0
-; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-FAKE16-NEXT:    s_and_b32 s4, 0xffff, s2
-; GFX12-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-FAKE16-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s4
-; GFX12-FAKE16-NEXT:    s_pack_hl_b32_b16 s2, s2, 0
-; GFX12-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
-; GFX12-FAKE16-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, v1
-; GFX12-FAKE16-NEXT:    s_pack_hl_b32_b16 s2, s3, 0
-; GFX12-FAKE16-NEXT:    s_and_b32 s3, 0xffff, s3
-; GFX12-FAKE16-NEXT:    global_store_b128 v1, v[0:3], s[0:1]
-; GFX12-FAKE16-NEXT:    s_wait_alu 0xfffe
-; GFX12-FAKE16-NEXT:    v_mov_b32_e32 v0, s3
-; GFX12-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
-; GFX12-FAKE16-NEXT:    global_store_b128 v1, v[0:3], s[0:1] offset:16
-; GFX12-FAKE16-NEXT:    s_endpgm
+; GFX12-LABEL: constant_zextload_v4i16_to_v4i64:
+; GFX12:       ; %bb.0:
+; GFX12-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
+; GFX12-NEXT:    s_wait_kmcnt 0x0
+; GFX12-NEXT:    s_load_b64 s[2:3], s[2:3], 0x0
+; GFX12-NEXT:    s_wait_kmcnt 0x0
+; GFX12-NEXT:    s_and_b32 s4, 0xffff, s2
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s4
+; GFX12-NEXT:    s_pack_hl_b32_b16 s2, s2, 0
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, v1
+; GFX12-NEXT:    s_pack_hl_b32_b16 s2, s3, 0
+; GFX12-NEXT:    s_and_b32 s3, 0xffff, s3
+; GFX12-NEXT:    global_store_b128 v1, v[0:3], s[0:1]
+; GFX12-NEXT:    s_wait_alu 0xfffe
+; GFX12-NEXT:    v_mov_b32_e32 v0, s3
+; GFX12-NEXT:    v_mov_b32_e32 v2, s2
+; GFX12-NEXT:    global_store_b128 v1, v[0:3], s[0:1] offset:16
+; GFX12-NEXT:    s_endpgm
   %load = load <4 x i16>, ptr addrspace(4) %in
   %ext = zext <4 x i16> %load to <4 x i64>
   store <4 x i64> %ext, ptr addrspace(1) %out
diff --git a/llvm/test/CodeGen/AMDGPU/load-constant-i8.ll b/llvm/test/CodeGen/AMDGPU/load-constant-i8.ll
index 70bdebedf85bc..b1bdfa667f57e 100644
--- a/llvm/test/CodeGen/AMDGPU/load-constant-i8.ll
+++ b/llvm/test/CodeGen/AMDGPU/load-constant-i8.ll
@@ -9969,46 +9969,25 @@ define amdgpu_kernel void @constant_sextload_v4i8_to_v4i16(ptr addrspace(1) %out
 ; EG-NEXT:     BFE_INT * T5.Y, PV.W, 0.0, literal.x,
 ; EG-NEXT:    8(1.121039e-44), 0(0.000000e+00)
 ;
-; GFX12-TRUE16-LABEL: constant_sextload_v4i8_to_v4i16:
-; GFX12-TRUE16:       ; %bb.0:
-; GFX12-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    s_load_b32 s2, s[2:3], 0x0
-; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    s_lshr_b32 s3, s2, 16
-; GFX12-TRUE16-NEXT:    s_ashr_i32 s4, s2, 24
-; GFX12-TRUE16-NEXT:    s_mov_b32 s5, s2
-; GFX12-TRUE16-NEXT:    s_sext_i32_i16 s2, s2
-; GFX12-TRUE16-NEXT:    s_bfe_i32 s5, s5, 0x80000
-; GFX12-TRUE16-NEXT:    s_lshr_b32 s2, s2, 8
-; GFX12-TRUE16-NEXT:    s_bfe_i32 s3, s3, 0x80000
-; GFX12-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s5, s2
-; GFX12-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v1, s3
-; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v0, s2
-; GFX12-TRUE16-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
-; GFX12-TRUE16-NEXT:    s_endpgm
-;
-; GFX12-FAKE16-LABEL: constant_sextload_v4i8_to_v4i16:
-; GFX12-FAKE16:       ; %bb.0:
-; GFX12-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-FAKE16-NEXT:    s_load_b32 s2, s[2:3], 0x0
-; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-FAKE16-NEXT:    s_lshr_b32 s3, s2, 16
-; GFX12-FAKE16-NEXT:    s_sext_i32_i16 s5, s2
-; GFX12-FAKE16-NEXT:    s_ashr_i32 s4, s2, 24
-; GFX12-FAKE16-NEXT:    s_bfe_i32 s2, s2, 0x80000
-; GFX12-FAKE16-NEXT:    s_lshr_b32 s5, s5, 8
-; GFX12-FAKE16-NEXT:    s_bfe_i32 s3, s3, 0x80000
-; GFX12-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s2, s5
-; GFX12-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
-; GFX12-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-FAKE16-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v1, s3
-; GFX12-FAKE16-NEXT:    v_mov_b32_e32 v0, s2
-; GFX12-FAKE16-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
-; GFX12-FAKE16-NEXT:    s_endpgm
+; GFX12-LABEL: constant_sextload_v4i8_to_v4i16:
+; GFX12:       ; %bb.0:
+; GFX12-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
+; GFX12-NEXT:    s_wait_kmcnt 0x0
+; GFX12-NEXT:    s_load_b32 s2, s[2:3], 0x0
+; GFX12-NEXT:    s_wait_kmcnt 0x0
+; GFX12-NEXT:    s_lshr_b32 s3, s2, 16
+; GFX12-NEXT:    s_sext_i32_i16 s5, s2
+; GFX12-NEXT:    s_ashr_i32 s4, s2, 24
+; GFX12-NEXT:    s_bfe_i32 s2, s2, 0x80000
+; GFX12-NEXT:    s_lshr_b32 s5, s5, 8
+; GFX12-NEXT:    s_bfe_i32 s3, s3, 0x80000
+; GFX12-NEXT:    s_pack_ll_b32_b16 s2, s2, s5
+; GFX12-NEXT:    s_pack_ll_b32_b16 s3, s3, s4
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v1, s3
+; GFX12-NEXT:    v_mov_b32_e32 v0, s2
+; GFX12-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
+; GFX12-NEXT:    s_endpgm
   %load = load <4 x i8>, ptr addrspace(4) %in
   %ext = sext <4 x i8> %load to <4 x i16>
   store <4 x i16> %ext, ptr addrspace(1) %out
@@ -10299,64 +10278,34 @@ define amdgpu_kernel void @constant_sextload_v8i8_to_v8i16(ptr addrspace(1) %out
 ; EG-NEXT:     BFE_INT * T6.Y, PS, 0.0, literal.y,
 ; EG-NEXT:    2(2.802597e-45), 8(1.121039e-44)
 ;
-; GFX12-TRUE16-LABEL: constant_sextload_v8i8_to_v8i16:
-; GFX12-TRUE16:       ; %bb.0:
-; GFX12-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    s_load_b64 s[2:3], s[2:3], 0x0
-; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-TRUE16-NEXT:    s_lshr_b32 s4, s2, 16
-; GFX12-TRUE16-NEXT:    s_lshr_b32 s5, s3, 16
-; GFX12-TRUE16-NEXT:    s_ashr_i32 s6, s3, 24
-; GFX12-TRUE16-NEXT:    s_bfe_i32 s7, s3, 0x80000
-; GFX12-TRUE16-NEXT:    s_sext_i32_i16 s3, s3
-; GFX12-TRUE16-NEXT:    s_ashr_i32 s8, s2, 24
-; GFX12-TRUE16-NEXT:    s_mov_b32 s9, s2
-; GFX12-TRUE16-NEXT:    s_sext_i32_i16 s2, s2
-; GFX12-TRUE16-NEXT:    s_bfe_i32 s5, s5, 0x80000
-; GFX12-TRUE16-NEXT:    s_lshr_b32 s3, s3, 8
-; GFX12-TRUE16-NEXT:    s_bfe_i32 s4, s4, 0x80000
-; GFX12-TRUE16-NEXT:    s_bfe_i32 s9, s9, 0x80000
-; GFX12-TRUE16-NEXT:    s_lshr_b32 s2, s2, 8
-; GFX12-TRUE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX12-TRUE16-NEXT:    s_pack_ll_b32_b16 s3, s7, s3
-; GFX12-TRUE16-NEXT:    s_pack_ll_b32_b16 s2, s9, s2
-; GFX12-TRUE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s8
-; GFX12-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v1, s4
-; GFX12-TRUE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v3, s5
-; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
-; GFX12-TRUE16-NEXT:    global_store_b128 v4, v[0:3], s[0:1]
-; GFX12-TRUE16-NEXT:    s_endpgm
-;
-; GFX12-FAKE16-LABEL: constant_sextload_v8i8_to_v8i16:
-; GFX12-FAKE16:       ; %bb.0:
-; GFX12-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
-; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-FAKE16-NEXT:    s_load_b64 s[2:3], s[2:3], 0x0
-; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
-; GFX12-FAKE16-NEXT:    s_lshr_b32 s4, s2, 16
-; GFX12-FAKE16-NEXT:    s_lshr_b32 s5, s3, 16
-; GFX12-FAKE16-NEXT:    s_ashr_i32 s6, s3, 24
-; GFX12-FAKE16-NEXT:    s_bfe_i32 s7, s3, 0x80000
-; GFX12-FAKE16-NEXT:    s_sext_i32_i16 s3, s3
-; GFX12-FAKE16-NEXT:    s_ashr_i32 s8, s2, 24
-; GFX12-FAKE16-NEXT:    s_bfe_i32 s9, s2, 0x80000
-; GFX12-FAKE16-NEXT:    s_sext_i32_i16 s2, s2
-; GFX12-FAKE16-NEXT:    s_bfe_i32 s5, s5, 0x80000
-; GFX12-FAKE16-NEXT:    s_lshr_b32 s3, s3, 8
-; GFX12-FAKE16-NEXT:    s_bfe_i32 s4, s4, 0x80000
-; GFX12-FAKE16-NEXT:    s_lshr_b32 s2, s2, 8
-; GFX12-FAKE16-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
-; GFX12-FAKE16-NEXT:    s_pack_ll_b32_b16 s3, s7, s3
-; GFX12-FAKE16-NEXT:    s_pack_ll_b32_b16 s2, s9, s2
-; GFX12-FAKE16-NEXT:    s_pack_ll_b32_b16 s4, s4, s8
-; GFX12-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX12-FAKE16-NEXT:    v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v1, s4
-; GFX12-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v3, s5
-; GFX12-FAKE16-NEXT:    v_mov_b32_e32 v2, s3
-; GFX12-FAKE16-NEXT:    global_store_b128 v4, v[0:3], s[0:1]
-; GFX12-FAKE16-NEXT:    s_endpgm
+; GFX12-LABEL: constant_sextload_v8i8_to_v8i16:
+; GFX12:       ; %bb.0:
+; GFX12-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
+; GFX12-NEXT:    s_wait_kmcnt 0x0
+; GFX12-NEXT:    s_load_b64 s[2:3], s[2:3], 0x0
+; GFX12-NEXT:    s_wait_kmcnt 0x0
+; GFX12-NEXT:    s_lshr_b32 s4, s2, 16
+; GFX12-NEXT:    s_lshr_b32 s5, s3, 16
+; GFX12-NEXT:    s_ashr_i32 s6, s3, 24
+; GFX12-NEXT:    s_bfe_i32 s7, s3, 0x80000
+; GFX12-NEXT:    s_sext_i32_i16 s3, s3
+; GFX12-NEXT:    s_ashr_i32 s8, s2, 24
+; GFX12-NEXT:    s_bfe_i32 s9, s2, 0x80000
+; GFX12-NEXT:    s_sext_i32_i16 s2, s2
+; GFX12-NEXT:    s_bfe_i32 s5, s5, 0x80000
+; GFX12-NEXT:    s_lshr_b32 s3, s3, 8
+; GFX12-NEXT:    s_bfe_i32 s4, s4, 0x80000
+; GFX12-NEXT:    s_lshr_b32 s2, s2, 8
+; GFX12-NEXT:    s_pack_ll_b32_b16 s5, s5, s6
+; GFX12-NEXT:    s_pack_ll_b32_b16 s3, s7, s3
+; GFX12-NEXT:    s_pack_ll_b32_b16 s2, s9, s2
+; GFX12-NEXT:    s_pack_ll_b32_b16 s4, s4, s8
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT:    v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v1, s4
+; GFX12-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v3, s5
+; GFX12-NEXT:    v_mov_b32_e32 v2, s3
+; GFX12-NEXT:    global_store_b128 v4, v[0:3], s[0:1]
+; GFX12-NEXT:    s_endpgm
   %load = load <8 x i8>, ptr addrspace(4) %in
   %ext = sext <8 x i8> %load to <8 x i16>
   store <8 x i16> %ext, ptr addrspace(1) %out
diff --git a/llvm/test/CodeGen/AMDGPU/select.f16.ll b/llvm/test/CodeGen/AMDGPU/select.f16.ll
index 51de691e0eccc..3f921ad850966 100644
--- a/llvm/test/CodeGen/AMDGPU/select.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/select.f16.ll
@@ -852,23 +852,23 @@ define amdgpu_kernel void @select_v2f16(
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s3, 0x31016000
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s22, s2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s23, s3
+; GFX11-TRUE16-NEXT:    s_mov_b32 s26, s2
+; GFX11-TRUE16-NEXT:    s_mov_b32 s27, s3
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s18, s2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s19, s3
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s6, s2
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s7, s3
-; GFX11-TRUE16-NEXT:    s_mov_b32 s26, s2
-; GFX11-TRUE16-NEXT:    s_mov_b32 s27, s3
 ; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s20, s12
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s21, s13
+; GFX11-TRUE16-NEXT:    s_mov_b32 s24, s14
+; GFX11-TRUE16-NEXT:    s_mov_b32 s25, s15
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s10
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s11
 ; GFX11-TRUE16-NEXT:    buffer_load_b32 v0, off, s[20:23], 0
 ; GFX11-TRUE16-NEXT:    buffer_load_b32 v1, off, s[16:19], 0
-; GFX11-TRUE16-NEXT:    s_mov_b32 s24, s14
-; GFX11-TRUE16-NEXT:    s_mov_b32 s25, s15
-; GFX11-TRUE16-NEXT:    buffer_load_b32 v2, off, s[4:7], 0
-; GFX11-TRUE16-NEXT:    buffer_load_b32 v3, off, s[24:27], 0
+; GFX11-TRUE16-NEXT:    buffer_load_b32 v2, off, s[24:27], 0
+; GFX11-TRUE16-NEXT:    buffer_load_b32 v3, off, s[4:7], 0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s9
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v0
@@ -876,17 +876,18 @@ define amdgpu_kernel void @select_v2f16(
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v1
 ; GFX11-TRUE16-NEXT:    v_cmp_lt_f16_e32 vcc_lo, v1.l, v0.l
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v2.l
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v2
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v3
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
 ; GFX11-TRUE16-NEXT:    v_cmp_lt_f16_e64 s0, v5.l, v4.l
-; GFX11-TRUE16-NEXT:    v_cndmask_b16 v2.l, v2.l, v3.l, vcc_lo
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v0.l, v1.l, s0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v2
+; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v3.l, v0.l, vcc_lo
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v2.l, v1.l, s0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s8
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; GFX11-TRUE16-NEXT:    buffer_store_b32 v0, off, s[0:3], 0
 ; GFX11-TRUE16-NEXT:    s_endpgm
 ;
@@ -1048,28 +1049,30 @@ define amdgpu_kernel void @select_v2f16_imm_a(
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s6
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s7
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s8
-; GFX11-TRUE16-NEXT:    buffer_load_b32 v0, off, s[12:15], 0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s9
+; GFX11-TRUE16-NEXT:    buffer_load_b32 v0, off, s[12:15], 0
+; GFX11-TRUE16-NEXT:    buffer_load_b32 v1, off, s[16:19], 0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s20, s10
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s21, s11
-; GFX11-TRUE16-NEXT:    buffer_load_b32 v1, off, s[16:19], 0
-; GFX11-TRUE16-NEXT:    buffer_load_b32 v2, off, s[20:23], 0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s5
+; GFX11-TRUE16-NEXT:    buffer_load_b32 v2, off, s[20:23], 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
 ; GFX11-TRUE16-NEXT:    v_cmp_lt_f16_e32 vcc_lo, 0.5, v0.l
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.l
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
 ; GFX11-TRUE16-NEXT:    v_cmp_lt_f16_e64 s0, 0x3900, v3.l
-; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v2.l, v1.l, vcc_lo
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v4.l
+; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v4.l, v0.l, s0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v5.l, v1.l, s0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s4
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; GFX11-TRUE16-NEXT:    buffer_store_b32 v0, off, s[0:3], 0
 ; GFX11-TRUE16-NEXT:    s_endpgm
 ;
@@ -1224,28 +1227,30 @@ define amdgpu_kernel void @select_v2f16_imm_b(
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s6
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s7
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s16, s8
-; GFX11-TRUE16-NEXT:    buffer_load_b32 v0, off, s[12:15], 0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s17, s9
+; GFX11-TRUE16-NEXT:    buffer_load_b32 v0, off, s[12:15], 0
+; GFX11-TRUE16-NEXT:    buffer_load_b32 v1, off, s[16:19], 0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s20, s10
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s21, s11
-; GFX11-TRUE16-NEXT:    buffer_load_b32 v1, off, s[16:19], 0
-; GFX11-TRUE16-NEXT:    buffer_load_b32 v2, off, s[20:23], 0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s1, s5
+; GFX11-TRUE16-NEXT:    buffer_load_b32 v2, off, s[20:23], 0
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
 ; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
 ; GFX11-TRUE16-NEXT:    v_cmp_gt_f16_e32 vcc_lo, 0.5, v0.l
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v1.l
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
+; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
 ; GFX11-TRUE16-NEXT:    v_cmp_gt_f16_e64 s0, 0x3900, v3.l
-; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v2.l, v1.l, vcc_lo
+; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v4.l
+; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, v4.l, v0.l, s0
-; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT:    v_cndmask_b16 v1.l, v5.l, v1.l, s0
+; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
 ; GFX11-TRUE16-NEXT:    s_mov_b32 s0, s4
 ; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
 ; GFX11-TRUE16-NEXT:    buffer_store_b32 v0, off, s[0:3], 0
 ; GFX11-TRUE16-NEXT:    s_endpgm
 ;



More information about the llvm-commits mailing list