[llvm] 1f08d3b - [AMDGPU] Further reduce attaching of implicit operands to spills

Jeffrey Byrnes via llvm-commits llvm-commits at lists.llvm.org
Thu Jan 19 14:31:53 PST 2023


Author: Jeffrey Byrnes
Date: 2023-01-19T14:31:07-08:00
New Revision: 1f08d3bc3a9a018534e52a47027dfb05cb4f55b3

URL: https://github.com/llvm/llvm-project/commit/1f08d3bc3a9a018534e52a47027dfb05cb4f55b3
DIFF: https://github.com/llvm/llvm-project/commit/1f08d3bc3a9a018534e52a47027dfb05cb4f55b3.diff

LOG: [AMDGPU] Further reduce attaching of implicit operands to spills

Extension of https://reviews.llvm.org/D141101 to even further reduce the amount of implicit operands we attach. The main benefit is to improve cability of post-ra scheduler, and reduce unneeded dependency resolution (e.g. inserting snops).

Unfortunately, we run into regressions if we completely minimize the amount implicit operands (naively), we run into some regressions (e.g. dual_movs are replaced with multiple calls to v_mov). This is even more reason to switch to LiveRegUnits.

Nonetheless, this patch removes the operands which we can for free (more or less).

Change-Id: Ib4f409202b36bdbc59eed615bc2d19fa8bd8c057

Differential Revision: https://reviews.llvm.org/D141557

Change-Id: I8b039e3c0d39436b384083f8beb947ee1b1730b2

Added: 
    

Modified: 
    llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
    llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
    llvm/test/CodeGen/AMDGPU/accvgpr-spill-scc-clobber.mir
    llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers.ll
    llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.ll
    llvm/test/CodeGen/AMDGPU/pei-build-av-spill.mir
    llvm/test/CodeGen/AMDGPU/pei-build-spill-partial-agpr.mir
    llvm/test/CodeGen/AMDGPU/pei-build-spill.mir
    llvm/test/CodeGen/AMDGPU/same-slot-agpr-sgpr.mir
    llvm/test/CodeGen/AMDGPU/sgpr-spill-to-vmem-scc-clobber.mir
    llvm/test/CodeGen/AMDGPU/sgpr-spill.mir
    llvm/test/CodeGen/AMDGPU/spill-reg-tuple-super-reg-use.mir
    llvm/test/CodeGen/AMDGPU/spill-special-sgpr.mir
    llvm/test/CodeGen/AMDGPU/spill192.mir
    llvm/test/CodeGen/AMDGPU/spill224.mir
    llvm/test/CodeGen/AMDGPU/spill288.mir
    llvm/test/CodeGen/AMDGPU/spill320.mir
    llvm/test/CodeGen/AMDGPU/spill352.mir
    llvm/test/CodeGen/AMDGPU/spill384.mir
    llvm/test/CodeGen/AMDGPU/vgpr-spill-scc-clobber.mir
    llvm/test/CodeGen/AMDGPU/vgpr-spill.mir

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 03de65fcd72f9..76347a3d230e3 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -1009,10 +1009,11 @@ void SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
     else
       SubIdx = SubIndices[SubIndices.size() - Idx - 1];
 
+    bool IsFirstSubreg = Idx == 0;
     bool UseKill = CanKillSuperReg && Idx == SubIndices.size() - 1;
 
     if (Opcode == AMDGPU::INSTRUCTION_LIST_END) {
-      Register ImpDefSuper = Idx == 0 ? Register(DestReg) : Register();
+      Register ImpDefSuper = IsFirstSubreg ? Register(DestReg) : Register();
       Register ImpUseSuper = SrcReg;
       indirectCopyToAGPR(*this, MBB, MI, DL, RI.getSubReg(DestReg, SubIdx),
                          RI.getSubReg(SrcReg, SubIdx), UseKill, *RS, Overlap,
@@ -1032,13 +1033,13 @@ void SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
         .addImm(0) // neg_hi
         .addImm(0) // clamp
         .addReg(SrcReg, getKillRegState(UseKill) | RegState::Implicit);
-      if (Idx == 0)
+      if (IsFirstSubreg)
         MIB.addReg(DestReg, RegState::Define | RegState::Implicit);
     } else {
       MachineInstrBuilder Builder =
         BuildMI(MBB, MI, DL, get(Opcode), RI.getSubReg(DestReg, SubIdx))
         .addReg(RI.getSubReg(SrcReg, SubIdx));
-      if (Idx == 0)
+      if (IsFirstSubreg)
         Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
 
       Builder.addReg(SrcReg, getKillRegState(UseKill) | RegState::Implicit);

diff  --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
index a11d9b4bf2fca..f7ce581f97363 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -1556,7 +1556,7 @@ void SIRegisterInfo::buildSpillLoadStore(
       auto MIB = spillVGPRtoAGPR(ST, MBB, MI, Index, Lane, Sub, IsKill);
       if (!MIB.getInstr())
         break;
-      if (NeedSuperRegDef || (IsSubReg && IsStore && Lane == LaneS && !i)) {
+      if (NeedSuperRegDef || (IsSubReg && IsStore && Lane == LaneS && IsFirstSubReg)) {
         MIB.addReg(ValueReg, RegState::ImplicitDefine);
         NeedSuperRegDef = false;
       }
@@ -1658,7 +1658,7 @@ void SIRegisterInfo::buildSpillLoadStore(
       MIB->setAsmPrinterFlag(MachineInstr::ReloadReuse);
     }
 
-    if (NeedSuperRegImpOperand)
+    if (NeedSuperRegImpOperand && (IsFirstSubReg || IsLastSubReg))
       MIB.addReg(ValueReg, RegState::Implicit | SrcDstRegState);
   }
 
@@ -1728,7 +1728,10 @@ bool SIRegisterInfo::spillSGPR(MachineBasicBlock::iterator MI, int Index,
               : Register(getSubReg(SB.SuperReg, SB.SplitParts[i]));
       SpilledReg Spill = VGPRSpills[i];
 
-      bool UseKill = SB.IsKill && i == SB.NumSubRegs - 1;
+      bool IsFirstSubreg = i == 0;
+      bool IsLastSubreg = i == SB.NumSubRegs - 1;
+      bool UseKill = SB.IsKill && IsLastSubreg;
+
 
       // Mark the "old value of vgpr" input undef only if this is the first sgpr
       // spill to this specific vgpr in the first basic block.
@@ -1738,19 +1741,19 @@ bool SIRegisterInfo::spillSGPR(MachineBasicBlock::iterator MI, int Index,
                      .addImm(Spill.Lane)
                      .addReg(Spill.VGPR);
       if (Indexes) {
-        if (i == 0)
+        if (IsFirstSubreg)
           Indexes->replaceMachineInstrInMaps(*MI, *MIB);
         else
           Indexes->insertMachineInstrInMaps(*MIB);
       }
 
-      if (i == 0 && SB.NumSubRegs > 1) {
+      if (IsFirstSubreg && SB.NumSubRegs > 1) {
         // We may be spilling a super-register which is only partially defined,
         // and need to ensure later spills think the value is defined.
         MIB.addReg(SB.SuperReg, RegState::ImplicitDefine);
       }
 
-      if (SB.NumSubRegs > 1)
+      if (SB.NumSubRegs > 1 && (IsFirstSubreg || IsLastSubreg))
         MIB.addReg(SB.SuperReg, getKillRegState(UseKill) | RegState::Implicit);
 
       // FIXME: Since this spills to another register instead of an actual

diff  --git a/llvm/test/CodeGen/AMDGPU/accvgpr-spill-scc-clobber.mir b/llvm/test/CodeGen/AMDGPU/accvgpr-spill-scc-clobber.mir
index 32453a600c309..53540e4a00049 100644
--- a/llvm/test/CodeGen/AMDGPU/accvgpr-spill-scc-clobber.mir
+++ b/llvm/test/CodeGen/AMDGPU/accvgpr-spill-scc-clobber.mir
@@ -2063,7 +2063,7 @@ body:             |
   ; GFX908-NEXT:   $vgpr0 = BUFFER_LOAD_DWORD_OFFEN $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
   ; GFX908-NEXT:   $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
   ; GFX908-NEXT:   $vgpr0 = BUFFER_LOAD_DWORD_OFFEN $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.1 + 4, addrspace 5)
-  ; GFX908-NEXT:   $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
+  ; GFX908-NEXT:   $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
   ; GFX908-NEXT:   $vgpr0 = BUFFER_LOAD_DWORD_OFFEN $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.1 + 8, addrspace 5)
   ; GFX908-NEXT:   $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
   ; GFX908-NEXT:   $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
@@ -2311,7 +2311,7 @@ body:             |
   ; GFX90A-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 704, 0, 0, implicit $exec :: (store (s32) into %stack.226, addrspace 5)
   ; GFX90A-NEXT:   $vgpr40 = V_MOV_B32_e32 8904, implicit $exec
   ; GFX90A-NEXT:   $agpr0 = BUFFER_LOAD_DWORD_OFFEN $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2 :: (load (s32) from %stack.1, addrspace 5)
-  ; GFX90A-NEXT:   $agpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2 :: (load (s32) from %stack.1 + 4, addrspace 5)
+  ; GFX90A-NEXT:   $agpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.1 + 4, addrspace 5)
   ; GFX90A-NEXT:   $agpr2 = BUFFER_LOAD_DWORD_OFFEN killed $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2 :: (load (s32) from %stack.1 + 8, addrspace 5)
   ; GFX90A-NEXT:   $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 704, 0, 0, implicit $exec :: (load (s32) from %stack.226, addrspace 5)
   ; GFX90A-NEXT:   S_CBRANCH_SCC1 %bb.2, implicit $scc
@@ -2562,7 +2562,7 @@ body:             |
   ; GFX908-FLATSCR-NEXT:   $vgpr0 = SCRATCH_LOAD_DWORD $vgpr1, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.1, addrspace 5)
   ; GFX908-FLATSCR-NEXT:   $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
   ; GFX908-FLATSCR-NEXT:   $vgpr0 = SCRATCH_LOAD_DWORD $vgpr1, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.1 + 4, addrspace 5)
-  ; GFX908-FLATSCR-NEXT:   $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
+  ; GFX908-FLATSCR-NEXT:   $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
   ; GFX908-FLATSCR-NEXT:   $vgpr0 = SCRATCH_LOAD_DWORD $vgpr1, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.1 + 8, addrspace 5)
   ; GFX908-FLATSCR-NEXT:   $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
   ; GFX908-FLATSCR-NEXT:   $vgpr1 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.2, addrspace 5)
@@ -5120,7 +5120,7 @@ body:             |
   ; GFX908-NEXT:   $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
   ; GFX908-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr0, $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.1, addrspace 5)
   ; GFX908-NEXT:   $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec
-  ; GFX908-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr0, $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.1 + 4, addrspace 5)
+  ; GFX908-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr0, $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.1 + 4, addrspace 5)
   ; GFX908-NEXT:   $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec
   ; GFX908-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr0, $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.1 + 8, addrspace 5)
   ; GFX908-NEXT:   $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
@@ -5368,7 +5368,7 @@ body:             |
   ; GFX90A-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 704, 0, 0, implicit $exec :: (store (s32) into %stack.226, addrspace 5)
   ; GFX90A-NEXT:   $vgpr40 = V_MOV_B32_e32 8904, implicit $exec
   ; GFX90A-NEXT:   BUFFER_STORE_DWORD_OFFEN $agpr0, $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.1, addrspace 5)
-  ; GFX90A-NEXT:   BUFFER_STORE_DWORD_OFFEN $agpr1, $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.1 + 4, addrspace 5)
+  ; GFX90A-NEXT:   BUFFER_STORE_DWORD_OFFEN $agpr1, $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.1 + 4, addrspace 5)
   ; GFX90A-NEXT:   BUFFER_STORE_DWORD_OFFEN $agpr2, killed $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.1 + 8, addrspace 5)
   ; GFX90A-NEXT:   $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 704, 0, 0, implicit $exec :: (load (s32) from %stack.226, addrspace 5)
   ; GFX90A-NEXT:   S_CBRANCH_SCC1 %bb.2, implicit $scc
@@ -5619,7 +5619,7 @@ body:             |
   ; GFX908-FLATSCR-NEXT:   $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
   ; GFX908-FLATSCR-NEXT:   SCRATCH_STORE_DWORD $vgpr0, $vgpr1, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.1, addrspace 5)
   ; GFX908-FLATSCR-NEXT:   $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec
-  ; GFX908-FLATSCR-NEXT:   SCRATCH_STORE_DWORD $vgpr0, $vgpr1, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.1 + 4, addrspace 5)
+  ; GFX908-FLATSCR-NEXT:   SCRATCH_STORE_DWORD $vgpr0, $vgpr1, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.1 + 4, addrspace 5)
   ; GFX908-FLATSCR-NEXT:   $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec
   ; GFX908-FLATSCR-NEXT:   SCRATCH_STORE_DWORD $vgpr0, $vgpr1, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.1 + 8, addrspace 5)
   ; GFX908-FLATSCR-NEXT:   $vgpr1 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.2, addrspace 5)

diff  --git a/llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers.ll b/llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers.ll
index 26b8b63ad5d6d..0a1c7eb5e5a69 100644
--- a/llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers.ll
+++ b/llvm/test/CodeGen/AMDGPU/agpr-copy-no-free-registers.ll
@@ -111,6 +111,10 @@ define void @no_free_vgprs_at_agpr_to_agpr_copy(float %v0, float %v1) #0 {
 ; GFX908-NEXT:    s_waitcnt vmcnt(0)
 ; GFX908-NEXT:    v_accvgpr_write_b32 a0, v39 ; Reload Reuse
 ; GFX908-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GFX908-NEXT:    v_accvgpr_write_b32 a11, v38 ; Reload Reuse
+; GFX908-NEXT:    v_accvgpr_write_b32 a12, v37 ; Reload Reuse
+; GFX908-NEXT:    v_accvgpr_write_b32 a13, v36 ; Reload Reuse
+; GFX908-NEXT:    v_accvgpr_write_b32 a14, v35 ; Reload Reuse
 ; GFX908-NEXT:    s_waitcnt vmcnt(0)
 ; GFX908-NEXT:    v_accvgpr_write_b32 a1, v39 ; Reload Reuse
 ; GFX908-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
@@ -140,10 +144,6 @@ define void @no_free_vgprs_at_agpr_to_agpr_copy(float %v0, float %v1) #0 {
 ; GFX908-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
 ; GFX908-NEXT:    s_waitcnt vmcnt(0)
 ; GFX908-NEXT:    v_accvgpr_write_b32 a10, v39 ; Reload Reuse
-; GFX908-NEXT:    v_accvgpr_write_b32 a11, v38 ; Reload Reuse
-; GFX908-NEXT:    v_accvgpr_write_b32 a12, v37 ; Reload Reuse
-; GFX908-NEXT:    v_accvgpr_write_b32 a13, v36 ; Reload Reuse
-; GFX908-NEXT:    v_accvgpr_write_b32 a14, v35 ; Reload Reuse
 ; GFX908-NEXT:    v_accvgpr_write_b32 a15, v34 ; Reload Reuse
 ; GFX908-NEXT:    ;;#ASMSTART
 ; GFX908-NEXT:    ; copy
@@ -208,28 +208,21 @@ define void @no_free_vgprs_at_agpr_to_agpr_copy(float %v0, float %v1) #0 {
 ; GFX90A-NEXT:    buffer_load_dword a0, off, s[0:3], s32 ; 4-byte Folded Reload
 ; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a4, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a5, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a6, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a7, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a8, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a9, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_waitcnt vmcnt(0)
+; GFX90A-NEXT:    s_waitcnt vmcnt(9)
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a10, v39 ; Reload Reuse
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a11, v38 ; Reload Reuse
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a12, v37 ; Reload Reuse
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a13, v36 ; Reload Reuse
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a14, v35 ; Reload Reuse
+; GFX90A-NEXT:    s_waitcnt vmcnt(0)
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a15, v34 ; Reload Reuse
 ; GFX90A-NEXT:    ;;#ASMSTART
 ; GFX90A-NEXT:    ; copy
@@ -1021,6 +1014,10 @@ define void @no_free_vgprs_at_sgpr_to_agpr_copy(float %v0, float %v1) #0 {
 ; GFX908-NEXT:    s_waitcnt vmcnt(0)
 ; GFX908-NEXT:    v_accvgpr_write_b32 a0, v39 ; Reload Reuse
 ; GFX908-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GFX908-NEXT:    v_accvgpr_write_b32 a11, v38 ; Reload Reuse
+; GFX908-NEXT:    v_accvgpr_write_b32 a12, v37 ; Reload Reuse
+; GFX908-NEXT:    v_accvgpr_write_b32 a13, v36 ; Reload Reuse
+; GFX908-NEXT:    v_accvgpr_write_b32 a14, v35 ; Reload Reuse
 ; GFX908-NEXT:    s_waitcnt vmcnt(0)
 ; GFX908-NEXT:    v_accvgpr_write_b32 a1, v39 ; Reload Reuse
 ; GFX908-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
@@ -1050,10 +1047,6 @@ define void @no_free_vgprs_at_sgpr_to_agpr_copy(float %v0, float %v1) #0 {
 ; GFX908-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
 ; GFX908-NEXT:    s_waitcnt vmcnt(0)
 ; GFX908-NEXT:    v_accvgpr_write_b32 a10, v39 ; Reload Reuse
-; GFX908-NEXT:    v_accvgpr_write_b32 a11, v38 ; Reload Reuse
-; GFX908-NEXT:    v_accvgpr_write_b32 a12, v37 ; Reload Reuse
-; GFX908-NEXT:    v_accvgpr_write_b32 a13, v36 ; Reload Reuse
-; GFX908-NEXT:    v_accvgpr_write_b32 a14, v35 ; Reload Reuse
 ; GFX908-NEXT:    v_accvgpr_write_b32 a15, v34 ; Reload Reuse
 ; GFX908-NEXT:    ;;#ASMSTART
 ; GFX908-NEXT:    ; copy
@@ -1119,29 +1112,21 @@ define void @no_free_vgprs_at_sgpr_to_agpr_copy(float %v0, float %v1) #0 {
 ; GFX90A-NEXT:    buffer_load_dword a0, off, s[0:3], s32 ; 4-byte Folded Reload
 ; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a4, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a5, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a6, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a7, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a8, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a9, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_nop 0
 ; GFX90A-NEXT:    buffer_load_dword a10, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GFX90A-NEXT:    s_waitcnt vmcnt(0)
+; GFX90A-NEXT:    s_waitcnt vmcnt(10)
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a11, v39 ; Reload Reuse
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a12, v38 ; Reload Reuse
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a13, v37 ; Reload Reuse
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a14, v36 ; Reload Reuse
+; GFX90A-NEXT:    s_waitcnt vmcnt(0)
 ; GFX90A-NEXT:    v_accvgpr_write_b32 a15, v35 ; Reload Reuse
 ; GFX90A-NEXT:    ;;#ASMSTART
 ; GFX90A-NEXT:    ; copy

diff  --git a/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.ll b/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.ll
index ecb949c46b8f7..eb1b4f0fa7d06 100644
--- a/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.ll
+++ b/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.ll
@@ -176,7 +176,6 @@ define float @mubuf_vgpr(<4 x i32> %i, i32 %c) #0 {
 ; W64-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
 ; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
 ; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
@@ -560,7 +559,6 @@ define void @mubuf_vgpr_adjacent_in_block(<4 x i32> %i, <4 x i32> %j, i32 %c, pt
 ; W64-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
 ; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
 ; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
@@ -614,7 +612,6 @@ define void @mubuf_vgpr_adjacent_in_block(<4 x i32> %i, <4 x i32> %j, i32 %c, pt
 ; W64-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
 ; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
 ; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
@@ -1063,7 +1060,6 @@ define void @mubuf_vgpr_outside_entry(<4 x i32> %i, <4 x i32> %j, i32 %c, ptr ad
 ; W64-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
 ; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
 ; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
@@ -1138,7 +1134,6 @@ define void @mubuf_vgpr_outside_entry(<4 x i32> %i, <4 x i32> %j, i32 %c, ptr ad
 ; W64-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
 ; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
 ; W64-O0-NEXT:    s_nop 0
 ; W64-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload

diff  --git a/llvm/test/CodeGen/AMDGPU/pei-build-av-spill.mir b/llvm/test/CodeGen/AMDGPU/pei-build-av-spill.mir
index 4f313b78b21aa..a9580da5af1da 100644
--- a/llvm/test/CodeGen/AMDGPU/pei-build-av-spill.mir
+++ b/llvm/test/CodeGen/AMDGPU/pei-build-av-spill.mir
@@ -166,10 +166,10 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_av_v3
     ; MUBUF: $vgpr0_vgpr1_vgpr2 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_av_v3
@@ -202,10 +202,10 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_v3
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_v3
@@ -256,12 +256,12 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_av_v4
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_av_v4
@@ -298,12 +298,12 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_v4
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_v4
@@ -358,14 +358,14 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_av_v5
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_av_v5
@@ -408,14 +408,14 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_v5
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_v5
@@ -476,16 +476,16 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_av_v6
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_av_v6
@@ -532,16 +532,16 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_v6
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_v6
@@ -606,18 +606,18 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_av_v7
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_av_v7
@@ -668,18 +668,18 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_v7
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_v7
@@ -748,20 +748,20 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_av_v8
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_av_v8
@@ -816,20 +816,20 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_v8
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_v8
@@ -902,36 +902,36 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_av_v16
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
     ; MUBUF-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 60, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_av_v16
@@ -974,12 +974,12 @@ body:             |
     ; FLATSCR-LABEL: name: test_spill_av_v16
     ; FLATSCR: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = IMPLICIT_DEF
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-NEXT: S_ENDPGM 0
     ; FLATSCR-V2A-LABEL: name: test_spill_av_v16
@@ -1022,36 +1022,36 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_v16
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 60, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_v16
@@ -1094,12 +1094,12 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_av_v16
     ; FLATSCR-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_av_v16
@@ -1160,68 +1160,68 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_av_v32
     ; MUBUF: $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 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, 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, implicit $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 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 56, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 60, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 64, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 68, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 72, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 76, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 80, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 84, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 88, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 92, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 96, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 100, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 104, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 108, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 112, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 116, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 120, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 60, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 64, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 68, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 72, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 76, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 80, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 84, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 88, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 92, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 96, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 100, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 104, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 108, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 112, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 116, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 120, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr31, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, implicit killed $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 :: (store (s32) into %stack.0 + 124, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; MUBUF-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 60, addrspace 5)
-    ; MUBUF-NEXT: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 64, addrspace 5)
-    ; MUBUF-NEXT: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 68, addrspace 5)
-    ; MUBUF-NEXT: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 72, addrspace 5)
-    ; MUBUF-NEXT: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 76, addrspace 5)
-    ; MUBUF-NEXT: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 80, addrspace 5)
-    ; MUBUF-NEXT: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 84, addrspace 5)
-    ; MUBUF-NEXT: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 88, addrspace 5)
-    ; MUBUF-NEXT: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 92, addrspace 5)
-    ; MUBUF-NEXT: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 96, addrspace 5)
-    ; MUBUF-NEXT: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 100, addrspace 5)
-    ; MUBUF-NEXT: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 104, addrspace 5)
-    ; MUBUF-NEXT: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 108, addrspace 5)
-    ; MUBUF-NEXT: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 112, addrspace 5)
-    ; MUBUF-NEXT: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 116, addrspace 5)
-    ; MUBUF-NEXT: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 120, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 60, addrspace 5)
+    ; MUBUF-NEXT: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 64, addrspace 5)
+    ; MUBUF-NEXT: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 68, addrspace 5)
+    ; MUBUF-NEXT: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 72, addrspace 5)
+    ; MUBUF-NEXT: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 76, addrspace 5)
+    ; MUBUF-NEXT: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 80, addrspace 5)
+    ; MUBUF-NEXT: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 84, addrspace 5)
+    ; MUBUF-NEXT: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 88, addrspace 5)
+    ; MUBUF-NEXT: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 92, addrspace 5)
+    ; MUBUF-NEXT: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 96, addrspace 5)
+    ; MUBUF-NEXT: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 100, addrspace 5)
+    ; MUBUF-NEXT: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 104, addrspace 5)
+    ; MUBUF-NEXT: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 108, addrspace 5)
+    ; MUBUF-NEXT: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 112, addrspace 5)
+    ; MUBUF-NEXT: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 116, addrspace 5)
+    ; MUBUF-NEXT: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 120, addrspace 5)
     ; MUBUF-NEXT: $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 124, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_av_v32
@@ -1296,20 +1296,20 @@ body:             |
     ; FLATSCR-LABEL: name: test_spill_av_v32
     ; FLATSCR: $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 = IMPLICIT_DEF
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, 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, implicit $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 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr16_vgpr17_vgpr18_vgpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr20_vgpr21_vgpr22_vgpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr24_vgpr25_vgpr26_vgpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr16_vgpr17_vgpr18_vgpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr20_vgpr21_vgpr22_vgpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr24_vgpr25_vgpr26_vgpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr28_vgpr29_vgpr30_vgpr31, $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, implicit killed $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 :: (store (s128) into %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr16_vgpr17_vgpr18_vgpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr20_vgpr21_vgpr22_vgpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr24_vgpr25_vgpr26_vgpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr16_vgpr17_vgpr18_vgpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr20_vgpr21_vgpr22_vgpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr24_vgpr25_vgpr26_vgpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr28_vgpr29_vgpr30_vgpr31 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-NEXT: S_ENDPGM 0
     ; FLATSCR-V2A-LABEL: name: test_spill_av_v32
@@ -1384,68 +1384,68 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_v32
     ; MUBUF-GFX90A: $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 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, 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, implicit $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 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 56, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 60, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 64, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 68, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 72, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 76, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 80, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 84, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 88, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 92, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 96, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 100, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 104, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 108, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 112, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 116, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 120, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 60, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 64, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 68, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 72, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 76, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 80, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 84, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 88, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 92, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 96, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 100, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 104, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 108, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 112, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 116, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 120, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr31, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, implicit killed $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 :: (store (s32) into %stack.0 + 124, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 60, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 64, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 68, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 72, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 76, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 80, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 84, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 88, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 92, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 96, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 100, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 104, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 108, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 112, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 116, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 120, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 60, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 64, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 68, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 72, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 76, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 80, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 84, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 88, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 92, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 96, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 100, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 104, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 108, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 112, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 116, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 120, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 124, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_v32
@@ -1520,20 +1520,20 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_av_v32
     ; FLATSCR-GFX90A: $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 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, 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, implicit $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 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr16_vgpr17_vgpr18_vgpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr20_vgpr21_vgpr22_vgpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr24_vgpr25_vgpr26_vgpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr16_vgpr17_vgpr18_vgpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr20_vgpr21_vgpr22_vgpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr24_vgpr25_vgpr26_vgpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr28_vgpr29_vgpr30_vgpr31, $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, implicit killed $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 :: (store (s128) into %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr16_vgpr17_vgpr18_vgpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr20_vgpr21_vgpr22_vgpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr24_vgpr25_vgpr26_vgpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr16_vgpr17_vgpr18_vgpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr20_vgpr21_vgpr22_vgpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr24_vgpr25_vgpr26_vgpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $vgpr28_vgpr29_vgpr30_vgpr31 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_av_v32
@@ -1786,13 +1786,13 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -1812,13 +1812,13 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -1836,10 +1836,10 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a3
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2 :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2 :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a3
@@ -1892,17 +1892,17 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -1924,17 +1924,17 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -1954,12 +1954,12 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a4
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3 :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3 :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a4
@@ -2016,21 +2016,21 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -2054,21 +2054,21 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -2090,14 +2090,14 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a5
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4 :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4 :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a5
@@ -2160,25 +2160,25 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -2204,25 +2204,25 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -2246,16 +2246,16 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a6
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a6
@@ -2322,29 +2322,29 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -2372,29 +2372,29 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -2420,18 +2420,18 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a7
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6 :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a7
@@ -2502,33 +2502,33 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -2558,33 +2558,33 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -2612,20 +2612,20 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a8
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a8
@@ -2700,37 +2700,37 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
     ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -2762,37 +2762,37 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
+    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 32, addrspace 5)
     ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -2822,22 +2822,22 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a9
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s32) from %stack.0 + 32, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a9
@@ -2866,10 +2866,10 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_av_a9
     ; FLATSCR-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr0_agpr1_agpr2_agpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORD_SADDR killed $agpr8, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr0_agpr1_agpr2_agpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr8 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8 :: (load (s32) from %stack.0 + 32, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_av_a9
@@ -2918,41 +2918,41 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
     ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -2986,41 +2986,41 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
+    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 36, addrspace 5)
     ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -3052,24 +3052,24 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a10
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s32) from %stack.0 + 36, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a10
@@ -3100,10 +3100,10 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_av_a10
     ; FLATSCR-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr0_agpr1_agpr2_agpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX2_SADDR killed $agpr8_agpr9, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (store (s64) into %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr0_agpr1_agpr2_agpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9 = SCRATCH_LOAD_DWORDX2_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9 :: (load (s64) from %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_av_a10
@@ -3154,45 +3154,45 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
     ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -3228,45 +3228,45 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
+    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 40, addrspace 5)
     ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -3300,26 +3300,26 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a11
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s32) from %stack.0 + 40, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a11
@@ -3352,10 +3352,10 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_av_a11
     ; FLATSCR-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr0_agpr1_agpr2_agpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX3_SADDR killed $agpr8_agpr9_agpr10, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (store (s96) into %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr0_agpr1_agpr2_agpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9_agpr10 = SCRATCH_LOAD_DWORDX3_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10 :: (load (s96) from %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_av_a11
@@ -3408,49 +3408,49 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr11, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
     ; MUBUF-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -3488,49 +3488,49 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr11, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 44, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
+    ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 44, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 44, addrspace 5)
     ; FLATSCR-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -3566,28 +3566,28 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a12
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s32) from %stack.0 + 44, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a12
@@ -3622,10 +3622,10 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_av_a12
     ; FLATSCR-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr0_agpr1_agpr2_agpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr8_agpr9_agpr10_agpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr0_agpr1_agpr2_agpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9_agpr10_agpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_av_a12
@@ -3680,65 +3680,65 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr11, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr12, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr13, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr14, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr15, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; MUBUF-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 60, addrspace 5)
     ; MUBUF-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -3784,65 +3784,65 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr11, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 44, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 44, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr12, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 48, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr13, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 52, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 52, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 52, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr14, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 56, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 56, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr15, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 60, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 44, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; FLATSCR-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; FLATSCR-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 52, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; FLATSCR-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 56, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; FLATSCR-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 60, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 60, addrspace 5)
     ; FLATSCR-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -3886,36 +3886,36 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a16
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 60, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a16
@@ -3958,12 +3958,12 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_av_a16
     ; FLATSCR-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr0_agpr1_agpr2_agpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr8_agpr9_agpr10_agpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr8_agpr9_agpr10_agpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr12_agpr13_agpr14_agpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr0_agpr1_agpr2_agpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9_agpr10_agpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9_agpr10_agpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr12_agpr13_agpr14_agpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_av_a16
@@ -4026,129 +4026,129 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr11, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr12, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr13, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr14, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr15, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 60, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr16, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 64, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 64, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr17, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 68, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 68, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr18, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 72, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 72, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr19, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 76, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 76, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr20, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 80, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 80, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr21, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 84, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 84, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr22, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 88, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 88, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr23, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 92, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 92, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr24, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 96, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 96, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr25, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 100, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 100, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr26, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 104, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 104, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr27, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 108, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 108, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr28, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 112, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 112, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr29, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 116, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 116, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr30, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 120, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 120, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr31, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 124, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; MUBUF-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 60, addrspace 5)
-    ; MUBUF-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 64, addrspace 5)
-    ; MUBUF-NEXT: $agpr16 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr16 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 68, addrspace 5)
-    ; MUBUF-NEXT: $agpr17 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr17 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 72, addrspace 5)
-    ; MUBUF-NEXT: $agpr18 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr18 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 76, addrspace 5)
-    ; MUBUF-NEXT: $agpr19 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr19 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 80, addrspace 5)
-    ; MUBUF-NEXT: $agpr20 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr20 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 84, addrspace 5)
-    ; MUBUF-NEXT: $agpr21 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr21 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 88, addrspace 5)
-    ; MUBUF-NEXT: $agpr22 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr22 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 92, addrspace 5)
-    ; MUBUF-NEXT: $agpr23 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr23 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 96, addrspace 5)
-    ; MUBUF-NEXT: $agpr24 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr24 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 100, addrspace 5)
-    ; MUBUF-NEXT: $agpr25 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr25 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 104, addrspace 5)
-    ; MUBUF-NEXT: $agpr26 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr26 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 108, addrspace 5)
-    ; MUBUF-NEXT: $agpr27 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr27 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 112, addrspace 5)
-    ; MUBUF-NEXT: $agpr28 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr28 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 116, addrspace 5)
-    ; MUBUF-NEXT: $agpr29 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr29 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 120, addrspace 5)
-    ; MUBUF-NEXT: $agpr30 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr30 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 124, addrspace 5)
     ; MUBUF-NEXT: $agpr31 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -4226,129 +4226,129 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr11, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 44, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 44, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr12, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 48, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr13, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 52, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 52, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 52, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr14, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 56, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 56, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr15, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 60, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 60, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 60, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr16, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 64, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 64, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr17, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 68, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 68, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 68, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 68, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr18, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 72, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 72, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 72, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 72, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr19, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 76, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 76, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 76, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 76, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr20, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 80, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 80, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr21, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 84, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 84, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 84, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 84, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr22, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 88, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 88, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 88, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 88, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr23, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 92, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 92, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 92, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 92, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr24, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 96, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 96, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr25, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 100, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 100, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 100, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 100, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr26, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 104, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 104, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 104, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 104, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr27, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 108, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 108, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 108, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 108, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr28, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 112, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 112, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 112, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr29, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 116, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 116, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 116, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 116, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr30, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 120, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 120, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 120, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 120, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr31, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 124, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 124, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 44, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; FLATSCR-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; FLATSCR-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 52, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; FLATSCR-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 56, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; FLATSCR-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 60, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 60, addrspace 5)
-    ; FLATSCR-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 64, addrspace 5)
-    ; FLATSCR-NEXT: $agpr16 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr16 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 68, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 68, addrspace 5)
-    ; FLATSCR-NEXT: $agpr17 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr17 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 72, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 72, addrspace 5)
-    ; FLATSCR-NEXT: $agpr18 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr18 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 76, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 76, addrspace 5)
-    ; FLATSCR-NEXT: $agpr19 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr19 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 80, addrspace 5)
-    ; FLATSCR-NEXT: $agpr20 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr20 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 84, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 84, addrspace 5)
-    ; FLATSCR-NEXT: $agpr21 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr21 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 88, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 88, addrspace 5)
-    ; FLATSCR-NEXT: $agpr22 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr22 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 92, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 92, addrspace 5)
-    ; FLATSCR-NEXT: $agpr23 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr23 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 96, addrspace 5)
-    ; FLATSCR-NEXT: $agpr24 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr24 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 100, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 100, addrspace 5)
-    ; FLATSCR-NEXT: $agpr25 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr25 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 104, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 104, addrspace 5)
-    ; FLATSCR-NEXT: $agpr26 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr26 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 108, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 108, addrspace 5)
-    ; FLATSCR-NEXT: $agpr27 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr27 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 112, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 112, addrspace 5)
-    ; FLATSCR-NEXT: $agpr28 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr28 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 116, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 116, addrspace 5)
-    ; FLATSCR-NEXT: $agpr29 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr29 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 120, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 120, addrspace 5)
-    ; FLATSCR-NEXT: $agpr30 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr30 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 124, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 124, addrspace 5)
     ; FLATSCR-NEXT: $agpr31 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -4424,68 +4424,68 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_av_a32
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 56, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 60, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 64, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 68, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 72, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 76, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 80, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 84, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 88, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 92, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 96, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 100, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 104, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 108, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 112, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 116, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 120, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 60, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 64, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 68, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 72, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 76, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 80, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 84, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 88, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 92, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 96, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 100, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 104, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 108, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 112, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 116, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 120, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr31, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 124, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 60, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 64, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 68, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 72, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 76, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 80, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 84, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 88, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 92, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 96, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 100, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 104, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 108, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 112, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 116, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 120, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 60, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 64, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 68, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 72, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 76, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 80, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 84, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 88, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 92, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 96, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 100, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 104, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 108, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 112, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 116, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 120, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 124, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_av_a32
@@ -4560,20 +4560,20 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_av_a32
     ; FLATSCR-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr0_agpr1_agpr2_agpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr8_agpr9_agpr10_agpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr12_agpr13_agpr14_agpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr16_agpr17_agpr18_agpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr20_agpr21_agpr22_agpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr24_agpr25_agpr26_agpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr8_agpr9_agpr10_agpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr12_agpr13_agpr14_agpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr16_agpr17_agpr18_agpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr20_agpr21_agpr22_agpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr24_agpr25_agpr26_agpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr28_agpr29_agpr30_agpr31, $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr0_agpr1_agpr2_agpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9_agpr10_agpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr12_agpr13_agpr14_agpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr16_agpr17_agpr18_agpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr20_agpr21_agpr22_agpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr24_agpr25_agpr26_agpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9_agpr10_agpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr12_agpr13_agpr14_agpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr16_agpr17_agpr18_agpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr20_agpr21_agpr22_agpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr24_agpr25_agpr26_agpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr28_agpr29_agpr30_agpr31 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_av_a32

diff  --git a/llvm/test/CodeGen/AMDGPU/pei-build-spill-partial-agpr.mir b/llvm/test/CodeGen/AMDGPU/pei-build-spill-partial-agpr.mir
index 19670e0c28547..ec50b6d548896 100644
--- a/llvm/test/CodeGen/AMDGPU/pei-build-spill-partial-agpr.mir
+++ b/llvm/test/CodeGen/AMDGPU/pei-build-spill-partial-agpr.mir
@@ -97,10 +97,10 @@ body:             |
     ; MUBUF-V2A-NEXT: {{  $}}
     ; MUBUF-V2A-NEXT: $vgpr0_vgpr1_vgpr2 = IMPLICIT_DEF
     ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-V2A-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr2, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2
     ; MUBUF-V2A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
     ; MUBUF-V2A-NEXT: $vgpr2 = V_ACCVGPR_READ_B32_e64 $agpr0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2
     ; MUBUF-V2A-NEXT: S_ENDPGM 0
     ; FLATSCR-V2A-LABEL: name: test_spill_v3_partial_agpr
@@ -179,12 +179,12 @@ body:             |
     ; MUBUF-V2A-NEXT: {{  $}}
     ; MUBUF-V2A-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 = IMPLICIT_DEF
     ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-V2A-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr2, implicit $exec
     ; MUBUF-V2A-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr3, implicit $exec
     ; MUBUF-V2A-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr4, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4
     ; MUBUF-V2A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
     ; MUBUF-V2A-NEXT: $vgpr2 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec
     ; MUBUF-V2A-NEXT: $vgpr3 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec
     ; MUBUF-V2A-NEXT: $vgpr4 = V_ACCVGPR_READ_B32_e64 $agpr0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4
@@ -277,17 +277,17 @@ body:             |
     ; MUBUF-V2A-NEXT: {{  $}}
     ; MUBUF-V2A-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = IMPLICIT_DEF
     ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-V2A-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr4, implicit $exec
     ; MUBUF-V2A-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr5, implicit $exec
     ; MUBUF-V2A-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr6, implicit $exec
     ; MUBUF-V2A-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr7, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
     ; MUBUF-V2A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-V2A-NEXT: $vgpr4 = V_ACCVGPR_READ_B32_e64 $agpr3, implicit $exec
     ; MUBUF-V2A-NEXT: $vgpr5 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec
     ; MUBUF-V2A-NEXT: $vgpr6 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec
@@ -331,32 +331,32 @@ body:             |
     ; MUBUF-V2A-NEXT: {{  $}}
     ; MUBUF-V2A-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = IMPLICIT_DEF
     ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-V2A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; MUBUF-V2A-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr11, implicit $exec
     ; MUBUF-V2A-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr12, implicit $exec
     ; MUBUF-V2A-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr13, implicit $exec
     ; MUBUF-V2A-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr14, implicit $exec
     ; MUBUF-V2A-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr15, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
     ; MUBUF-V2A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-V2A-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-V2A-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
     ; MUBUF-V2A-NEXT: $vgpr11 = V_ACCVGPR_READ_B32_e64 $agpr4, implicit $exec
     ; MUBUF-V2A-NEXT: $vgpr12 = V_ACCVGPR_READ_B32_e64 $agpr3, implicit $exec
     ; MUBUF-V2A-NEXT: $vgpr13 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec
@@ -368,17 +368,17 @@ body:             |
     ; FLATSCR-V2A-NEXT: {{  $}}
     ; FLATSCR-V2A-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = IMPLICIT_DEF
     ; FLATSCR-V2A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-V2A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-V2A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
     ; FLATSCR-V2A-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr11, implicit $exec
-    ; FLATSCR-V2A-NEXT: SCRATCH_STORE_DWORDX3_SADDR killed $vgpr8_vgpr9_vgpr10, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s96) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-V2A-NEXT: SCRATCH_STORE_DWORDX3_SADDR killed $vgpr8_vgpr9_vgpr10, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s96) into %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-V2A-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr15, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
     ; FLATSCR-V2A-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr14, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
     ; FLATSCR-V2A-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr13, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
     ; FLATSCR-V2A-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr12, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
     ; FLATSCR-V2A-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-V2A-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-V2A-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
     ; FLATSCR-V2A-NEXT: $vgpr11 = V_ACCVGPR_READ_B32_e64 $agpr4, implicit $exec
-    ; FLATSCR-V2A-NEXT: $vgpr8_vgpr9_vgpr10 = SCRATCH_LOAD_DWORDX3_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s96) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-V2A-NEXT: $vgpr8_vgpr9_vgpr10 = SCRATCH_LOAD_DWORDX3_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s96) from %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-V2A-NEXT: $vgpr15 = V_ACCVGPR_READ_B32_e64 $agpr0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
     ; FLATSCR-V2A-NEXT: $vgpr14 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
     ; FLATSCR-V2A-NEXT: $vgpr13 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15

diff  --git a/llvm/test/CodeGen/AMDGPU/pei-build-spill.mir b/llvm/test/CodeGen/AMDGPU/pei-build-spill.mir
index 6d222408c787f..bd052daff5cd4 100644
--- a/llvm/test/CodeGen/AMDGPU/pei-build-spill.mir
+++ b/llvm/test/CodeGen/AMDGPU/pei-build-spill.mir
@@ -166,10 +166,10 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_v3
     ; MUBUF: $vgpr0_vgpr1_vgpr2 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_v3
@@ -202,10 +202,10 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_v3
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_v3
@@ -256,12 +256,12 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_v4
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_v4
@@ -298,12 +298,12 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_v4
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3 :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_v4
@@ -358,14 +358,14 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_v5
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_v5
@@ -408,14 +408,14 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_v5
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_v5
@@ -476,16 +476,16 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_v6
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_v6
@@ -532,16 +532,16 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_v6
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_v6
@@ -606,20 +606,20 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_v8
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_v8
@@ -674,20 +674,20 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_v8
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_v8
@@ -760,36 +760,36 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_v16
     ; MUBUF: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
     ; MUBUF-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 60, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_v16
@@ -832,12 +832,12 @@ body:             |
     ; FLATSCR-LABEL: name: test_spill_v16
     ; FLATSCR: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = IMPLICIT_DEF
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-NEXT: S_ENDPGM 0
     ; FLATSCR-V2A-LABEL: name: test_spill_v16
@@ -880,36 +880,36 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_v16
     ; MUBUF-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s32) from %stack.0 + 60, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_v16
@@ -952,12 +952,12 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_v16
     ; FLATSCR-GFX90A: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_v16
@@ -1018,68 +1018,68 @@ body:             |
     ; MUBUF-LABEL: name: test_spill_v32
     ; MUBUF: $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 = IMPLICIT_DEF
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, 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, implicit $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 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 56, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 60, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 64, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 68, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 72, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 76, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 80, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 84, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 88, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 92, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 96, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 100, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 104, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 108, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 112, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 116, addrspace 5)
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 120, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 60, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 64, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 68, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 72, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 76, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 80, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 84, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 88, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 92, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 96, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 100, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 104, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 108, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 112, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 116, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 120, addrspace 5)
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr31, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, implicit killed $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 :: (store (s32) into %stack.0 + 124, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; MUBUF-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 60, addrspace 5)
-    ; MUBUF-NEXT: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 64, addrspace 5)
-    ; MUBUF-NEXT: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 68, addrspace 5)
-    ; MUBUF-NEXT: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 72, addrspace 5)
-    ; MUBUF-NEXT: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 76, addrspace 5)
-    ; MUBUF-NEXT: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 80, addrspace 5)
-    ; MUBUF-NEXT: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 84, addrspace 5)
-    ; MUBUF-NEXT: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 88, addrspace 5)
-    ; MUBUF-NEXT: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 92, addrspace 5)
-    ; MUBUF-NEXT: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 96, addrspace 5)
-    ; MUBUF-NEXT: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 100, addrspace 5)
-    ; MUBUF-NEXT: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 104, addrspace 5)
-    ; MUBUF-NEXT: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 108, addrspace 5)
-    ; MUBUF-NEXT: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 112, addrspace 5)
-    ; MUBUF-NEXT: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 116, addrspace 5)
-    ; MUBUF-NEXT: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 120, addrspace 5)
+    ; MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 60, addrspace 5)
+    ; MUBUF-NEXT: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 64, addrspace 5)
+    ; MUBUF-NEXT: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 68, addrspace 5)
+    ; MUBUF-NEXT: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 72, addrspace 5)
+    ; MUBUF-NEXT: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 76, addrspace 5)
+    ; MUBUF-NEXT: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 80, addrspace 5)
+    ; MUBUF-NEXT: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 84, addrspace 5)
+    ; MUBUF-NEXT: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 88, addrspace 5)
+    ; MUBUF-NEXT: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 92, addrspace 5)
+    ; MUBUF-NEXT: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 96, addrspace 5)
+    ; MUBUF-NEXT: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 100, addrspace 5)
+    ; MUBUF-NEXT: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 104, addrspace 5)
+    ; MUBUF-NEXT: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 108, addrspace 5)
+    ; MUBUF-NEXT: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 112, addrspace 5)
+    ; MUBUF-NEXT: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 116, addrspace 5)
+    ; MUBUF-NEXT: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 120, addrspace 5)
     ; MUBUF-NEXT: $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 124, addrspace 5)
     ; MUBUF-NEXT: S_ENDPGM 0
     ; MUBUF-V2A-LABEL: name: test_spill_v32
@@ -1154,20 +1154,20 @@ body:             |
     ; FLATSCR-LABEL: name: test_spill_v32
     ; FLATSCR: $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 = IMPLICIT_DEF
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, 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, implicit $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 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr16_vgpr17_vgpr18_vgpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr20_vgpr21_vgpr22_vgpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr24_vgpr25_vgpr26_vgpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr16_vgpr17_vgpr18_vgpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr20_vgpr21_vgpr22_vgpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr24_vgpr25_vgpr26_vgpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr28_vgpr29_vgpr30_vgpr31, $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, implicit killed $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 :: (store (s128) into %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr16_vgpr17_vgpr18_vgpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr20_vgpr21_vgpr22_vgpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-NEXT: $vgpr24_vgpr25_vgpr26_vgpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr16_vgpr17_vgpr18_vgpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr20_vgpr21_vgpr22_vgpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-NEXT: $vgpr24_vgpr25_vgpr26_vgpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr28_vgpr29_vgpr30_vgpr31 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-NEXT: S_ENDPGM 0
     ; FLATSCR-V2A-LABEL: name: test_spill_v32
@@ -1242,68 +1242,68 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_v32
     ; MUBUF-GFX90A: $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 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, 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, implicit $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 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 56, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 60, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 64, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 68, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 72, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 76, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 80, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 84, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 88, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 92, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 96, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 100, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 104, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 108, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 112, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 116, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, implicit $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 :: (store (s32) into %stack.0 + 120, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 60, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 64, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 68, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 72, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 76, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 80, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 84, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 88, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 92, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 96, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 100, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 104, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 108, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 112, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 116, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 120, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr31, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, implicit killed $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 :: (store (s32) into %stack.0 + 124, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 60, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 64, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 68, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 72, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 76, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 80, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 84, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 88, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 92, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 96, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 100, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 104, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 108, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 112, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 116, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 120, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 60, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 64, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 68, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 72, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 76, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 80, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 84, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 88, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 92, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 96, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 100, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 104, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 108, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 112, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 116, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 120, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, 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 :: (load (s32) from %stack.0 + 124, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_v32
@@ -1378,20 +1378,20 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_v32
     ; FLATSCR-GFX90A: $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 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, 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, implicit $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 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr16_vgpr17_vgpr18_vgpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr20_vgpr21_vgpr22_vgpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr24_vgpr25_vgpr26_vgpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, implicit $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 :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr4_vgpr5_vgpr6_vgpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr8_vgpr9_vgpr10_vgpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr12_vgpr13_vgpr14_vgpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr16_vgpr17_vgpr18_vgpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr20_vgpr21_vgpr22_vgpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr24_vgpr25_vgpr26_vgpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $vgpr28_vgpr29_vgpr30_vgpr31, $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, implicit killed $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 :: (store (s128) into %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $vgpr0_vgpr1_vgpr2_vgpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr16_vgpr17_vgpr18_vgpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr20_vgpr21_vgpr22_vgpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $vgpr24_vgpr25_vgpr26_vgpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr4_vgpr5_vgpr6_vgpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr8_vgpr9_vgpr10_vgpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr12_vgpr13_vgpr14_vgpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr16_vgpr17_vgpr18_vgpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr20_vgpr21_vgpr22_vgpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $vgpr24_vgpr25_vgpr26_vgpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $vgpr28_vgpr29_vgpr30_vgpr31 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, 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 :: (load (s128) from %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_v32
@@ -1644,13 +1644,13 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -1670,13 +1670,13 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -1694,10 +1694,10 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_a3
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2 :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2 :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2 :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_a3
@@ -1750,17 +1750,17 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -1782,17 +1782,17 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -1812,12 +1812,12 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_a4
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3 :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3 :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_a4
@@ -1874,21 +1874,21 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -1912,21 +1912,21 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -1948,14 +1948,14 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_a5
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4 :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4 :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4 :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_a5
@@ -2018,25 +2018,25 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -2062,25 +2062,25 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -2104,16 +2104,16 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_a6
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5 :: (load (s32) from %stack.0 + 20, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_a6
@@ -2180,33 +2180,33 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -2236,33 +2236,33 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -2290,20 +2290,20 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_a8
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7 :: (load (s32) from %stack.0 + 28, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_a8
@@ -2378,65 +2378,65 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr11, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr12, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr13, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr14, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr15, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; MUBUF-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; MUBUF-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 60, addrspace 5)
     ; MUBUF-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -2482,65 +2482,65 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr11, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 44, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 44, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr12, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 48, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr13, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 52, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 52, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 52, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr14, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 56, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 56, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr15, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 60, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 44, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; FLATSCR-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; FLATSCR-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 52, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; FLATSCR-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 56, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; FLATSCR-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
+    ; FLATSCR-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 60, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 60, addrspace 5)
     ; FLATSCR-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -2584,36 +2584,36 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_a16
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s32) from %stack.0 + 60, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_a16
@@ -2656,12 +2656,12 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_a16
     ; FLATSCR-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr0_agpr1_agpr2_agpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr8_agpr9_agpr10_agpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr8_agpr9_agpr10_agpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr12_agpr13_agpr14_agpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr0_agpr1_agpr2_agpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9_agpr10_agpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9_agpr10_agpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr12_agpr13_agpr14_agpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_a16
@@ -2724,129 +2724,129 @@ body:             |
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr11, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr12, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr13, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr14, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr15, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 60, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr16, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 64, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 64, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr17, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 68, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 68, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr18, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 72, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 72, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr19, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 76, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 76, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr20, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 80, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 80, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr21, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 84, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 84, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr22, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 88, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 88, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr23, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 92, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 92, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr24, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 96, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 96, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr25, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 100, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 100, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr26, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 104, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 104, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr27, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 108, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 108, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr28, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 112, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 112, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr29, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 116, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 116, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr30, implicit $exec
-    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 120, addrspace 5)
+    ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 120, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr31, implicit $exec
     ; MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 124, addrspace 5)
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; MUBUF-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; MUBUF-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 60, addrspace 5)
-    ; MUBUF-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 64, addrspace 5)
-    ; MUBUF-NEXT: $agpr16 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr16 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 68, addrspace 5)
-    ; MUBUF-NEXT: $agpr17 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr17 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 72, addrspace 5)
-    ; MUBUF-NEXT: $agpr18 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr18 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 76, addrspace 5)
-    ; MUBUF-NEXT: $agpr19 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr19 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 80, addrspace 5)
-    ; MUBUF-NEXT: $agpr20 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr20 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 84, addrspace 5)
-    ; MUBUF-NEXT: $agpr21 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr21 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 88, addrspace 5)
-    ; MUBUF-NEXT: $agpr22 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr22 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 92, addrspace 5)
-    ; MUBUF-NEXT: $agpr23 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr23 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 96, addrspace 5)
-    ; MUBUF-NEXT: $agpr24 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr24 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 100, addrspace 5)
-    ; MUBUF-NEXT: $agpr25 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr25 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 104, addrspace 5)
-    ; MUBUF-NEXT: $agpr26 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr26 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 108, addrspace 5)
-    ; MUBUF-NEXT: $agpr27 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr27 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 112, addrspace 5)
-    ; MUBUF-NEXT: $agpr28 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr28 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 116, addrspace 5)
-    ; MUBUF-NEXT: $agpr29 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr29 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 120, addrspace 5)
-    ; MUBUF-NEXT: $agpr30 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; MUBUF-NEXT: $agpr30 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 124, addrspace 5)
     ; MUBUF-NEXT: $agpr31 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; MUBUF-NEXT: S_ENDPGM 0
@@ -2924,129 +2924,129 @@ body:             |
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr1, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 4, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr2, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr3, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr4, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 16, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr5, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 20, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr6, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 24, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr7, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 28, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr8, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 32, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr9, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 36, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr10, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 40, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr11, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 44, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 44, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 44, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr12, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 48, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr13, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 52, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 52, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 52, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr14, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 56, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 56, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 56, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr15, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 60, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 60, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 60, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 60, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr16, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 64, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 64, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr17, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 68, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 68, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 68, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 68, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr18, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 72, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 72, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 72, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 72, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr19, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 76, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 76, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 76, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 76, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr20, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 80, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 80, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr21, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 84, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 84, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 84, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 84, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr22, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 88, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 88, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 88, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 88, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr23, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 92, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 92, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 92, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 92, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr24, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 96, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 96, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr25, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 100, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 100, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 100, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 100, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr26, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 104, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 104, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 104, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 104, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr27, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 108, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 108, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 108, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 108, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr28, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 112, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 112, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 112, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr29, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 116, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 116, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 116, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 116, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr30, implicit $exec
-    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 120, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 120, addrspace 5)
+    ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 120, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0 + 120, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = V_ACCVGPR_READ_B32_e64 killed $agpr31, implicit $exec
     ; FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr32, 124, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 124, addrspace 5)
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0, addrspace 5)
     ; FLATSCR-NEXT: $agpr0 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 4, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 8, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr2 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 12, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr3 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr4 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 20, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr5 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 24, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr6 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 28, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr7 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr8 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 36, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr9 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 40, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr10 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 44, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; FLATSCR-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr11 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; FLATSCR-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr12 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 52, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; FLATSCR-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr13 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 56, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; FLATSCR-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr14 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 60, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 60, addrspace 5)
-    ; FLATSCR-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr15 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 64, addrspace 5)
-    ; FLATSCR-NEXT: $agpr16 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr16 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 68, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 68, addrspace 5)
-    ; FLATSCR-NEXT: $agpr17 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr17 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 72, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 72, addrspace 5)
-    ; FLATSCR-NEXT: $agpr18 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr18 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 76, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 76, addrspace 5)
-    ; FLATSCR-NEXT: $agpr19 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr19 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 80, addrspace 5)
-    ; FLATSCR-NEXT: $agpr20 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr20 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 84, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 84, addrspace 5)
-    ; FLATSCR-NEXT: $agpr21 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr21 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 88, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 88, addrspace 5)
-    ; FLATSCR-NEXT: $agpr22 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr22 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 92, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 92, addrspace 5)
-    ; FLATSCR-NEXT: $agpr23 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr23 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 96, addrspace 5)
-    ; FLATSCR-NEXT: $agpr24 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr24 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 100, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 100, addrspace 5)
-    ; FLATSCR-NEXT: $agpr25 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr25 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 104, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 104, addrspace 5)
-    ; FLATSCR-NEXT: $agpr26 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr26 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 108, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 108, addrspace 5)
-    ; FLATSCR-NEXT: $agpr27 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr27 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 112, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 112, addrspace 5)
-    ; FLATSCR-NEXT: $agpr28 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr28 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 116, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 116, addrspace 5)
-    ; FLATSCR-NEXT: $agpr29 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr29 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 120, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 120, addrspace 5)
-    ; FLATSCR-NEXT: $agpr30 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
+    ; FLATSCR-NEXT: $agpr30 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr32, 124, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %stack.0 + 124, addrspace 5)
     ; FLATSCR-NEXT: $agpr31 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31
     ; FLATSCR-NEXT: S_ENDPGM 0
@@ -3122,68 +3122,68 @@ body:             |
     ; MUBUF-GFX90A-LABEL: name: test_spill_a32
     ; MUBUF-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 = IMPLICIT_DEF
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 56, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 60, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 64, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 68, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 72, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 76, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 80, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 84, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 88, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 92, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 96, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 100, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 104, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 108, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 112, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 116, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 120, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr4, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr5, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr6, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr7, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr8, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr9, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr10, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr11, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr12, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr13, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr14, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr15, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 60, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr16, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 64, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr17, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 68, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr18, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 72, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr19, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 76, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr20, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 80, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr21, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 84, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr22, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 88, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr23, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 92, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr24, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 96, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr25, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 100, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr26, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 104, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr27, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 108, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr28, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 112, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr29, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 116, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr30, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 120, addrspace 5)
     ; MUBUF-GFX90A-NEXT: BUFFER_STORE_DWORD_OFFSET killed $agpr31, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s32) into %stack.0 + 124, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 4, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 8, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 12, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 16, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 20, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 24, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 28, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 32, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 36, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 40, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 44, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 48, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 52, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 56, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 60, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 64, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 68, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 72, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 76, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 80, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 84, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 88, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 92, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 96, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 100, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 104, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 108, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 112, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 116, addrspace 5)
-    ; MUBUF-GFX90A-NEXT: $agpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 120, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 4, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 8, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 12, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 16, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 20, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 24, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 24, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 28, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 28, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 32, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 32, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 36, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 36, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 40, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 44, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 48, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 52, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 56, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 56, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 60, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 60, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 64, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 64, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 68, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 68, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 72, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 72, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 76, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 76, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 80, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 80, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 84, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 84, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 88, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 88, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 92, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 92, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 96, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 96, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 100, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 100, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 104, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 104, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 108, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 108, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 112, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 112, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 116, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 116, addrspace 5)
+    ; MUBUF-GFX90A-NEXT: $agpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 120, 0, 0, implicit $exec :: (load (s32) from %stack.0 + 120, addrspace 5)
     ; MUBUF-GFX90A-NEXT: $agpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 124, 0, 0, implicit $exec, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s32) from %stack.0 + 124, addrspace 5)
     ; MUBUF-GFX90A-NEXT: S_ENDPGM 0
     ; MUBUF-GFX90A-V2A-LABEL: name: test_spill_a32
@@ -3258,20 +3258,20 @@ body:             |
     ; FLATSCR-GFX90A-LABEL: name: test_spill_a32
     ; FLATSCR-GFX90A: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 = IMPLICIT_DEF
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr0_agpr1_agpr2_agpr3, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr8_agpr9_agpr10_agpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr12_agpr13_agpr14_agpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr16_agpr17_agpr18_agpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr20_agpr21_agpr22_agpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr24_agpr25_agpr26_agpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, implicit $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr4_agpr5_agpr6_agpr7, $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr8_agpr9_agpr10_agpr11, $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr12_agpr13_agpr14_agpr15, $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr16_agpr17_agpr18_agpr19, $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr20_agpr21_agpr22_agpr23, $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr24_agpr25_agpr26_agpr27, $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (store (s128) into %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: SCRATCH_STORE_DWORDX4_SADDR killed $agpr28_agpr29_agpr30_agpr31, $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, implicit killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (store (s128) into %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr0_agpr1_agpr2_agpr3 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 0, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9_agpr10_agpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr12_agpr13_agpr14_agpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr16_agpr17_agpr18_agpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr20_agpr21_agpr22_agpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
-    ; FLATSCR-GFX90A-NEXT: $agpr24_agpr25_agpr26_agpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr4_agpr5_agpr6_agpr7 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 16, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 16, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr8_agpr9_agpr10_agpr11 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 32, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 32, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr12_agpr13_agpr14_agpr15 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 48, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 48, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr16_agpr17_agpr18_agpr19 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 64, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 64, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr20_agpr21_agpr22_agpr23 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 80, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 80, align 4, addrspace 5)
+    ; FLATSCR-GFX90A-NEXT: $agpr24_agpr25_agpr26_agpr27 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 96, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %stack.0 + 96, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: $agpr28_agpr29_agpr30_agpr31 = SCRATCH_LOAD_DWORDX4_SADDR $sgpr32, 112, 0, implicit $exec, implicit $flat_scr, implicit-def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 :: (load (s128) from %stack.0 + 112, align 4, addrspace 5)
     ; FLATSCR-GFX90A-NEXT: S_ENDPGM 0
     ; FLATSCR-GFX90A-V2A-LABEL: name: test_spill_a32

diff  --git a/llvm/test/CodeGen/AMDGPU/same-slot-agpr-sgpr.mir b/llvm/test/CodeGen/AMDGPU/same-slot-agpr-sgpr.mir
index eaa6cab40467e..b323d880c9827 100644
--- a/llvm/test/CodeGen/AMDGPU/same-slot-agpr-sgpr.mir
+++ b/llvm/test/CodeGen/AMDGPU/same-slot-agpr-sgpr.mir
@@ -21,7 +21,7 @@ body: |
     ; CHECK-NEXT: $sgpr6_sgpr7 = S_MOV_B64 $exec
     ; CHECK-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr0
     ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
-    ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, undef $vgpr0, implicit $sgpr4_sgpr5
+    ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, undef $vgpr0
     ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0, implicit killed $sgpr4_sgpr5
     ; CHECK-NEXT: $agpr1 = V_ACCVGPR_WRITE_B32_e64 killed $vgpr0, implicit $exec
     ; CHECK-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
@@ -54,7 +54,7 @@ body: |
     ; CHECK-NEXT: $sgpr6_sgpr7 = S_MOV_B64 $exec
     ; CHECK-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr0
     ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
-    ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, undef $vgpr0, implicit $sgpr4_sgpr5
+    ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, undef $vgpr0
     ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0, implicit killed $sgpr4_sgpr5
     ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
     ; CHECK-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)

diff  --git a/llvm/test/CodeGen/AMDGPU/sgpr-spill-to-vmem-scc-clobber.mir b/llvm/test/CodeGen/AMDGPU/sgpr-spill-to-vmem-scc-clobber.mir
index 7798036f64bf0..8f7d2f7d05927 100644
--- a/llvm/test/CodeGen/AMDGPU/sgpr-spill-to-vmem-scc-clobber.mir
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-spill-to-vmem-scc-clobber.mir
@@ -77,7 +77,7 @@ body:             |
   ; VMEM-GFX8-NEXT:   $sgpr4_sgpr5 = S_MOV_B64 $exec
   ; VMEM-GFX8-NEXT:   $exec = S_MOV_B64 3, implicit-def $vgpr0
   ; VMEM-GFX8-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
-  ; VMEM-GFX8-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 0, undef $vgpr0, implicit $sgpr8_sgpr9
+  ; VMEM-GFX8-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 0, undef $vgpr0
   ; VMEM-GFX8-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 1, $vgpr0, implicit $sgpr8_sgpr9
   ; VMEM-GFX8-NEXT:   $vgpr1 = V_MOV_B32_e32 8200, implicit $exec
   ; VMEM-GFX8-NEXT:   BUFFER_STORE_DWORD_OFFEN killed $vgpr0, killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
@@ -333,7 +333,7 @@ body:             |
   ; VMEM-GFX8-NEXT:   $sgpr4_sgpr5 = S_MOV_B64 $exec
   ; VMEM-GFX8-NEXT:   $exec = S_MOV_B64 3
   ; VMEM-GFX8-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
-  ; VMEM-GFX8-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 0, undef $vgpr0, implicit $sgpr8_sgpr9
+  ; VMEM-GFX8-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 0, undef $vgpr0
   ; VMEM-GFX8-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 1, $vgpr0, implicit $sgpr8_sgpr9
   ; VMEM-GFX8-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.3, addrspace 5)
   ; VMEM-GFX8-NEXT:   $vgpr1 = V_MOV_B32_e32 8200, implicit $exec

diff  --git a/llvm/test/CodeGen/AMDGPU/sgpr-spill.mir b/llvm/test/CodeGen/AMDGPU/sgpr-spill.mir
index 8c0211ac25068..e8db1c8aa6ccb 100644
--- a/llvm/test/CodeGen/AMDGPU/sgpr-spill.mir
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-spill.mir
@@ -474,7 +474,7 @@ body:             |
     ; GCN64-MUBUF-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
     ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr2
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2, implicit $sgpr12_sgpr13
+    ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2
     ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr13, 1, $vgpr2, implicit killed $sgpr12_sgpr13
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 8, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
     ; GCN64-MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -483,7 +483,7 @@ body:             |
     ; GCN64-MUBUF-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
     ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr3
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF-NEXT: $vgpr3 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr3, implicit $sgpr12_sgpr13
+    ; GCN64-MUBUF-NEXT: $vgpr3 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr3
     ; GCN64-MUBUF-NEXT: $vgpr3 = V_WRITELANE_B32 $sgpr13, 1, $vgpr3, implicit $sgpr12_sgpr13
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 8, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
     ; GCN64-MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -492,8 +492,8 @@ body:             |
     ; GCN64-MUBUF-NEXT: $sgpr6_sgpr7 = S_MOV_B64 $exec
     ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 7, implicit-def $vgpr1
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1, implicit $sgpr12_sgpr13_sgpr14
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1, implicit $sgpr12_sgpr13_sgpr14
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1
     ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr14, 2, $vgpr1, implicit killed $sgpr12_sgpr13_sgpr14
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 16, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
     ; GCN64-MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -502,9 +502,9 @@ body:             |
     ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
     ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 15, implicit-def $vgpr0
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0
     ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 28, 0, 0, implicit $exec :: (store (s32) into %stack.3, addrspace 5)
     ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -513,10 +513,10 @@ body:             |
     ; GCN64-MUBUF-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
     ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 31, implicit-def $vgpr2
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr13, 1, $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr14, 2, $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr15, 3, $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2
+    ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr13, 1, $vgpr2
+    ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr14, 2, $vgpr2
+    ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr15, 3, $vgpr2
     ; GCN64-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr16, 4, $vgpr2, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 44, 0, 0, implicit $exec :: (store (s32) into %stack.4, addrspace 5)
     ; GCN64-MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -525,13 +525,13 @@ body:             |
     ; GCN64-MUBUF-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
     ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 255, implicit-def $vgpr1
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr14, 2, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr15, 3, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr16, 4, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr17, 5, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr18, 6, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr14, 2, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr15, 3, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr16, 4, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr17, 5, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr18, 6, $vgpr1
     ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr19, 7, $vgpr1, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 64, 0, 0, implicit $exec :: (store (s32) into %stack.5, addrspace 5)
     ; GCN64-MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -540,21 +540,21 @@ body:             |
     ; GCN64-MUBUF-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
     ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 65535, implicit-def $vgpr0
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0
+    ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0
     ; GCN64-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr27, 15, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 96, 0, 0, implicit $exec :: (store (s32) into %stack.6, addrspace 5)
     ; GCN64-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -563,37 +563,37 @@ body:             |
     ; GCN64-MUBUF-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
     ; GCN64-MUBUF-NEXT: $exec = S_MOV_B64 4294967295, implicit-def $vgpr1
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr65, 1, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr66, 2, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr67, 3, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr68, 4, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr69, 5, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr70, 6, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr71, 7, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr72, 8, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr73, 9, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr74, 10, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr75, 11, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr76, 12, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr77, 13, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr78, 14, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr79, 15, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr80, 16, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr81, 17, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr82, 18, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr83, 19, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr84, 20, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr85, 21, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr86, 22, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr87, 23, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr88, 24, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr89, 25, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr90, 26, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr91, 27, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr92, 28, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr93, 29, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr94, 30, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr65, 1, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr66, 2, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr67, 3, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr68, 4, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr69, 5, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr70, 6, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr71, 7, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr72, 8, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr73, 9, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr74, 10, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr75, 11, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr76, 12, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr77, 13, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr78, 14, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr79, 15, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr80, 16, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr81, 17, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr82, 18, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr83, 19, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr84, 20, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr85, 21, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr86, 22, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr87, 23, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr88, 24, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr89, 25, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr90, 26, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr91, 27, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr92, 28, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr93, 29, $vgpr1
+    ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr94, 30, $vgpr1
     ; GCN64-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr95, 31, $vgpr1, implicit killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
     ; GCN64-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 160, 0, 0, implicit $exec :: (store (s32) into %stack.7, addrspace 5)
     ; GCN64-MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr28_sgpr29_sgpr30_sgpr31, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -637,7 +637,7 @@ body:             |
     ; GCN32-MUBUF-NEXT: $sgpr2 = S_MOV_B32 $exec_lo
     ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 3, implicit-def $vgpr2
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2, implicit $sgpr12_sgpr13
+    ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2
     ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr13, 1, $vgpr2, implicit killed $sgpr12_sgpr13
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 8, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
     ; GCN32-MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -646,7 +646,7 @@ body:             |
     ; GCN32-MUBUF-NEXT: $sgpr1 = S_MOV_B32 $exec_lo
     ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 3, implicit-def $vgpr3
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF-NEXT: $vgpr3 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr3, implicit $sgpr12_sgpr13
+    ; GCN32-MUBUF-NEXT: $vgpr3 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr3
     ; GCN32-MUBUF-NEXT: $vgpr3 = V_WRITELANE_B32 $sgpr13, 1, $vgpr3, implicit $sgpr12_sgpr13
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 8, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
     ; GCN32-MUBUF-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -655,8 +655,8 @@ body:             |
     ; GCN32-MUBUF-NEXT: $sgpr3 = S_MOV_B32 $exec_lo
     ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 7, implicit-def $vgpr1
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1, implicit $sgpr12_sgpr13_sgpr14
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1, implicit $sgpr12_sgpr13_sgpr14
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1
     ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr14, 2, $vgpr1, implicit killed $sgpr12_sgpr13_sgpr14
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 16, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
     ; GCN32-MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -665,9 +665,9 @@ body:             |
     ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
     ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 15, implicit-def $vgpr0
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0
     ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 28, 0, 0, implicit $exec :: (store (s32) into %stack.3, addrspace 5)
     ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -676,10 +676,10 @@ body:             |
     ; GCN32-MUBUF-NEXT: $sgpr2 = S_MOV_B32 $exec_lo
     ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 31, implicit-def $vgpr2
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr13, 1, $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr14, 2, $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr15, 3, $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2
+    ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr13, 1, $vgpr2
+    ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr14, 2, $vgpr2
+    ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr15, 3, $vgpr2
     ; GCN32-MUBUF-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr16, 4, $vgpr2, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 44, 0, 0, implicit $exec :: (store (s32) into %stack.4, addrspace 5)
     ; GCN32-MUBUF-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -688,13 +688,13 @@ body:             |
     ; GCN32-MUBUF-NEXT: $sgpr1 = S_MOV_B32 $exec_lo
     ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 255, implicit-def $vgpr1
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr14, 2, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr15, 3, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr16, 4, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr17, 5, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr18, 6, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr14, 2, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr15, 3, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr16, 4, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr17, 5, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr18, 6, $vgpr1
     ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr19, 7, $vgpr1, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 64, 0, 0, implicit $exec :: (store (s32) into %stack.5, addrspace 5)
     ; GCN32-MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -703,21 +703,21 @@ body:             |
     ; GCN32-MUBUF-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
     ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 65535, implicit-def $vgpr0
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0
+    ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0
     ; GCN32-MUBUF-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr27, 15, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 96, 0, 0, implicit $exec :: (store (s32) into %stack.6, addrspace 5)
     ; GCN32-MUBUF-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -726,37 +726,37 @@ body:             |
     ; GCN32-MUBUF-NEXT: $sgpr2 = S_MOV_B32 $exec_lo
     ; GCN32-MUBUF-NEXT: $exec_lo = S_MOV_B32 4294967295, implicit-def $vgpr1
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr65, 1, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr66, 2, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr67, 3, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr68, 4, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr69, 5, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr70, 6, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr71, 7, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr72, 8, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr73, 9, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr74, 10, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr75, 11, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr76, 12, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr77, 13, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr78, 14, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr79, 15, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr80, 16, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr81, 17, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr82, 18, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr83, 19, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr84, 20, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr85, 21, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr86, 22, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr87, 23, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr88, 24, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr89, 25, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr90, 26, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr91, 27, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr92, 28, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr93, 29, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr94, 30, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr65, 1, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr66, 2, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr67, 3, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr68, 4, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr69, 5, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr70, 6, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr71, 7, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr72, 8, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr73, 9, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr74, 10, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr75, 11, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr76, 12, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr77, 13, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr78, 14, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr79, 15, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr80, 16, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr81, 17, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr82, 18, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr83, 19, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr84, 20, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr85, 21, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr86, 22, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr87, 23, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr88, 24, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr89, 25, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr90, 26, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr91, 27, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr92, 28, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr93, 29, $vgpr1
+    ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr94, 30, $vgpr1
     ; GCN32-MUBUF-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr95, 31, $vgpr1, implicit killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
     ; GCN32-MUBUF-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 160, 0, 0, implicit $exec :: (store (s32) into %stack.7, addrspace 5)
     ; GCN32-MUBUF-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -796,7 +796,7 @@ body:             |
     ; GCN64-FLATSCR-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
     ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr2
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr2, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2, implicit $sgpr12_sgpr13
+    ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2
     ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr13, 1, $vgpr2, implicit killed $sgpr12_sgpr13
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr2, $sgpr33, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.1, addrspace 5)
     ; GCN64-FLATSCR-NEXT: $vgpr2 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -805,7 +805,7 @@ body:             |
     ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
     ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr3
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr3, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR-NEXT: $vgpr3 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr3, implicit $sgpr12_sgpr13
+    ; GCN64-FLATSCR-NEXT: $vgpr3 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr3
     ; GCN64-FLATSCR-NEXT: $vgpr3 = V_WRITELANE_B32 $sgpr13, 1, $vgpr3, implicit $sgpr12_sgpr13
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr3, $sgpr33, 8, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.1, addrspace 5)
     ; GCN64-FLATSCR-NEXT: $vgpr3 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -814,8 +814,8 @@ body:             |
     ; GCN64-FLATSCR-NEXT: $sgpr6_sgpr7 = S_MOV_B64 $exec
     ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 7, implicit-def $vgpr1
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr1, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1, implicit $sgpr12_sgpr13_sgpr14
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1, implicit $sgpr12_sgpr13_sgpr14
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1
     ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr14, 2, $vgpr1, implicit killed $sgpr12_sgpr13_sgpr14
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr1, $sgpr33, 16, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.2, addrspace 5)
     ; GCN64-FLATSCR-NEXT: $vgpr1 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -824,9 +824,9 @@ body:             |
     ; GCN64-FLATSCR-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
     ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 15, implicit-def $vgpr0
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0
     ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 28, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.3, addrspace 5)
     ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -835,10 +835,10 @@ body:             |
     ; GCN64-FLATSCR-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
     ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 31, implicit-def $vgpr2
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr2, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr13, 1, $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr14, 2, $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
-    ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr15, 3, $vgpr2, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
+    ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr2
+    ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr13, 1, $vgpr2
+    ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr14, 2, $vgpr2
+    ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr15, 3, $vgpr2
     ; GCN64-FLATSCR-NEXT: $vgpr2 = V_WRITELANE_B32 $sgpr16, 4, $vgpr2, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr2, $sgpr33, 44, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.4, addrspace 5)
     ; GCN64-FLATSCR-NEXT: $vgpr2 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -847,13 +847,13 @@ body:             |
     ; GCN64-FLATSCR-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
     ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 255, implicit-def $vgpr1
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr1, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr14, 2, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr15, 3, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr16, 4, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr17, 5, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr18, 6, $vgpr1, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr13, 1, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr14, 2, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr15, 3, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr16, 4, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr17, 5, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr18, 6, $vgpr1
     ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr19, 7, $vgpr1, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr1, $sgpr33, 64, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.5, addrspace 5)
     ; GCN64-FLATSCR-NEXT: $vgpr1 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -862,21 +862,21 @@ body:             |
     ; GCN64-FLATSCR-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
     ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 65535, implicit-def $vgpr0
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
-    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0, implicit $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 0, undef $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 1, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 2, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr15, 3, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr16, 4, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr17, 5, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr18, 6, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr19, 7, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr20, 8, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr21, 9, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr22, 10, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr23, 11, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr24, 12, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr25, 13, $vgpr0
+    ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr26, 14, $vgpr0
     ; GCN64-FLATSCR-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr27, 15, $vgpr0, implicit killed $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr0, $sgpr33, 96, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.6, addrspace 5)
     ; GCN64-FLATSCR-NEXT: $vgpr0 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -885,37 +885,37 @@ body:             |
     ; GCN64-FLATSCR-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
     ; GCN64-FLATSCR-NEXT: $exec = S_MOV_B64 4294967295, implicit-def $vgpr1
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr1, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr65, 1, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr66, 2, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr67, 3, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr68, 4, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr69, 5, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr70, 6, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr71, 7, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr72, 8, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr73, 9, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr74, 10, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr75, 11, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr76, 12, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr77, 13, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr78, 14, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr79, 15, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr80, 16, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr81, 17, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr82, 18, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr83, 19, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr84, 20, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr85, 21, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr86, 22, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr87, 23, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr88, 24, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr89, 25, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr90, 26, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr91, 27, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr92, 28, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr93, 29, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
-    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr94, 30, $vgpr1, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr64, 0, undef $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr65, 1, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr66, 2, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr67, 3, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr68, 4, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr69, 5, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr70, 6, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr71, 7, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr72, 8, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr73, 9, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr74, 10, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr75, 11, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr76, 12, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr77, 13, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr78, 14, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr79, 15, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr80, 16, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr81, 17, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr82, 18, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr83, 19, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr84, 20, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr85, 21, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr86, 22, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr87, 23, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr88, 24, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr89, 25, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr90, 26, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr91, 27, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr92, 28, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr93, 29, $vgpr1
+    ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr94, 30, $vgpr1
     ; GCN64-FLATSCR-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr95, 31, $vgpr1, implicit killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
     ; GCN64-FLATSCR-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr1, $sgpr33, 160, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.7, addrspace 5)
     ; GCN64-FLATSCR-NEXT: $vgpr1 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)

diff  --git a/llvm/test/CodeGen/AMDGPU/spill-reg-tuple-super-reg-use.mir b/llvm/test/CodeGen/AMDGPU/spill-reg-tuple-super-reg-use.mir
index a4400670ab550..6c18983672d7d 100644
--- a/llvm/test/CodeGen/AMDGPU/spill-reg-tuple-super-reg-use.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill-reg-tuple-super-reg-use.mir
@@ -27,8 +27,8 @@ body:             |
     ; GCN-NEXT: $exec = S_MOV_B64 killed $sgpr8_sgpr9
     ; GCN-NEXT: renamable $sgpr1 = COPY $sgpr2
     ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr0, 0, $vgpr0, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr0_sgpr1_sgpr2_sgpr3
-    ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr1, 1, $vgpr0, implicit $sgpr0_sgpr1_sgpr2_sgpr3
-    ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr2, 2, $vgpr0, implicit $sgpr0_sgpr1_sgpr2_sgpr3
+    ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr1, 1, $vgpr0
+    ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr2, 2, $vgpr0
     ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr3, 3, $vgpr0, implicit $sgpr0_sgpr1_sgpr2_sgpr3
     ; GCN-NEXT: renamable $sgpr8 = COPY killed renamable $sgpr1
     ; GCN-NEXT: $sgpr0_sgpr1 = S_XOR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
@@ -64,8 +64,8 @@ body:             |
     ; GCN-NEXT: $exec = S_MOV_B64 killed $sgpr8_sgpr9
     ; GCN-NEXT: renamable $sgpr1 = COPY $sgpr2
     ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr0, 0, $vgpr0, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr0_sgpr1_sgpr2_sgpr3
-    ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr1, 1, $vgpr0, implicit $sgpr0_sgpr1_sgpr2_sgpr3
-    ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr2, 2, $vgpr0, implicit $sgpr0_sgpr1_sgpr2_sgpr3
+    ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr1, 1, $vgpr0
+    ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr2, 2, $vgpr0
     ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr3, 3, $vgpr0, implicit killed $sgpr0_sgpr1_sgpr2_sgpr3
     ; GCN-NEXT: $sgpr0_sgpr1 = S_XOR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
     ; GCN-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
@@ -95,8 +95,8 @@ body:             |
     ; GCN-NEXT: {{  $}}
     ; GCN-NEXT: renamable $vgpr1 = COPY $vgpr2
     ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr0, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0, addrspace 5)
-    ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr1, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr2, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr1, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr2, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr3, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; GCN-NEXT: renamable $vgpr8 = COPY killed renamable $vgpr1
     ; GCN-NEXT: S_ENDPGM 0, implicit $vgpr8
@@ -125,8 +125,8 @@ body:             |
     ; GCN-NEXT: {{  $}}
     ; GCN-NEXT: renamable $vgpr1 = COPY $vgpr2
     ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0, addrspace 5)
-    ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     ; GCN-NEXT: S_ENDPGM 0
     renamable $vgpr1 = COPY $vgpr2

diff  --git a/llvm/test/CodeGen/AMDGPU/spill-special-sgpr.mir b/llvm/test/CodeGen/AMDGPU/spill-special-sgpr.mir
index 3914c3b2fe073..a5703a553fcba 100644
--- a/llvm/test/CodeGen/AMDGPU/spill-special-sgpr.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill-special-sgpr.mir
@@ -51,7 +51,7 @@ body:             |
     ; GFX9-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
     ; GFX9-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr2
     ; GFX9-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr12_sgpr13_sgpr14_sgpr15, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GFX9-NEXT: $vgpr2 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr2, implicit $vcc
+    ; GFX9-NEXT: $vgpr2 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr2
     ; GFX9-NEXT: $vgpr2 = V_WRITELANE_B32 $vcc_hi, 1, $vgpr2, implicit $vcc
     ; GFX9-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr12_sgpr13_sgpr14_sgpr15, $sgpr33, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
     ; GFX9-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr12_sgpr13_sgpr14_sgpr15, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -60,7 +60,7 @@ body:             |
     ; GFX9-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
     ; GFX9-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr1
     ; GFX9-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr12_sgpr13_sgpr14_sgpr15, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GFX9-NEXT: $vgpr1 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr1, implicit $vcc
+    ; GFX9-NEXT: $vgpr1 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr1
     ; GFX9-NEXT: $vgpr1 = V_WRITELANE_B32 $vcc_hi, 1, $vgpr1, implicit killed $vcc
     ; GFX9-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr12_sgpr13_sgpr14_sgpr15, $sgpr33, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
     ; GFX9-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr12_sgpr13_sgpr14_sgpr15, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -87,7 +87,7 @@ body:             |
     ; GFX10-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
     ; GFX10-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr2
     ; GFX10-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GFX10-NEXT: $vgpr2 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr2, implicit $vcc
+    ; GFX10-NEXT: $vgpr2 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr2
     ; GFX10-NEXT: $vgpr2 = V_WRITELANE_B32 $vcc_hi, 1, $vgpr2, implicit $vcc
     ; GFX10-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
     ; GFX10-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -96,7 +96,7 @@ body:             |
     ; GFX10-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
     ; GFX10-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr1
     ; GFX10-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GFX10-NEXT: $vgpr1 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr1, implicit $vcc
+    ; GFX10-NEXT: $vgpr1 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr1
     ; GFX10-NEXT: $vgpr1 = V_WRITELANE_B32 $vcc_hi, 1, $vgpr1, implicit killed $vcc
     ; GFX10-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
     ; GFX10-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr96_sgpr97_sgpr98_sgpr99, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -117,7 +117,7 @@ body:             |
     ; GFX11-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
     ; GFX11-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr2
     ; GFX11-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr2, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GFX11-NEXT: $vgpr2 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr2, implicit $vcc
+    ; GFX11-NEXT: $vgpr2 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr2
     ; GFX11-NEXT: $vgpr2 = V_WRITELANE_B32 $vcc_hi, 1, $vgpr2, implicit $vcc
     ; GFX11-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr2, $sgpr33, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0, addrspace 5)
     ; GFX11-NEXT: $vgpr2 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)
@@ -126,7 +126,7 @@ body:             |
     ; GFX11-NEXT: $sgpr2_sgpr3 = S_MOV_B64 $exec
     ; GFX11-NEXT: $exec = S_MOV_B64 3, implicit-def $vgpr1
     ; GFX11-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr1, $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %fixed-stack.0, align 16, addrspace 5)
-    ; GFX11-NEXT: $vgpr1 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr1, implicit $vcc
+    ; GFX11-NEXT: $vgpr1 = V_WRITELANE_B32 $vcc_lo, 0, undef $vgpr1
     ; GFX11-NEXT: $vgpr1 = V_WRITELANE_B32 $vcc_hi, 1, $vgpr1, implicit killed $vcc
     ; GFX11-NEXT: SCRATCH_STORE_DWORD_SADDR killed $vgpr1, $sgpr33, 4, 0, implicit $exec, implicit $flat_scr :: (store (s32) into %stack.0, addrspace 5)
     ; GFX11-NEXT: $vgpr1 = SCRATCH_LOAD_DWORD_SADDR $sgpr33, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %fixed-stack.0, align 16, addrspace 5)

diff  --git a/llvm/test/CodeGen/AMDGPU/spill192.mir b/llvm/test/CodeGen/AMDGPU/spill192.mir
index df0c836b556e2..1834b889502a8 100644
--- a/llvm/test/CodeGen/AMDGPU/spill192.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill192.mir
@@ -36,10 +36,10 @@ body: |
   ; EXPANDED-NEXT: {{  $}}
   ; EXPANDED-NEXT:   S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 killed $sgpr9, 5, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
   ; EXPANDED-NEXT:   S_CBRANCH_SCC1 %bb.1, implicit undef $scc
   ; EXPANDED-NEXT: {{  $}}

diff  --git a/llvm/test/CodeGen/AMDGPU/spill224.mir b/llvm/test/CodeGen/AMDGPU/spill224.mir
index 09f208246995c..8a5ffd32d01a7 100644
--- a/llvm/test/CodeGen/AMDGPU/spill224.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill224.mir
@@ -34,11 +34,11 @@ body: |
   ; EXPANDED-NEXT: {{  $}}
   ; EXPANDED-NEXT:   S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 killed $sgpr10, 6, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
   ; EXPANDED-NEXT:   S_CBRANCH_SCC1 %bb.1, implicit undef $scc
   ; EXPANDED-NEXT: {{  $}}

diff  --git a/llvm/test/CodeGen/AMDGPU/spill288.mir b/llvm/test/CodeGen/AMDGPU/spill288.mir
index 6e6b7f8bd7e32..c5d56dd2cb04f 100644
--- a/llvm/test/CodeGen/AMDGPU/spill288.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill288.mir
@@ -34,13 +34,13 @@ body: |
   ; EXPANDED-NEXT: {{  $}}
   ; EXPANDED-NEXT:   S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 killed $sgpr12, 8, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
   ; EXPANDED-NEXT:   S_CBRANCH_SCC1 %bb.1, implicit undef $scc
   ; EXPANDED-NEXT: {{  $}}

diff  --git a/llvm/test/CodeGen/AMDGPU/spill320.mir b/llvm/test/CodeGen/AMDGPU/spill320.mir
index a5a1074d1e7dd..b54d43cae3acd 100644
--- a/llvm/test/CodeGen/AMDGPU/spill320.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill320.mir
@@ -34,14 +34,14 @@ body: |
   ; EXPANDED-NEXT: {{  $}}
   ; EXPANDED-NEXT:   S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr12, 8, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr12, 8, $vgpr0
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 killed $sgpr13, 9, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
   ; EXPANDED-NEXT:   S_CBRANCH_SCC1 %bb.1, implicit undef $scc
   ; EXPANDED-NEXT: {{  $}}

diff  --git a/llvm/test/CodeGen/AMDGPU/spill352.mir b/llvm/test/CodeGen/AMDGPU/spill352.mir
index 12a15152e6a10..0e215c707e1fb 100644
--- a/llvm/test/CodeGen/AMDGPU/spill352.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill352.mir
@@ -34,15 +34,15 @@ body: |
   ; EXPANDED-NEXT: {{  $}}
   ; EXPANDED-NEXT:   S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr12, 8, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr13, 9, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr12, 8, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr13, 9, $vgpr0
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 killed $sgpr14, 10, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
   ; EXPANDED-NEXT:   S_CBRANCH_SCC1 %bb.1, implicit undef $scc
   ; EXPANDED-NEXT: {{  $}}

diff  --git a/llvm/test/CodeGen/AMDGPU/spill384.mir b/llvm/test/CodeGen/AMDGPU/spill384.mir
index 6029ff52e9b5b..f0cee62a3baeb 100644
--- a/llvm/test/CodeGen/AMDGPU/spill384.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill384.mir
@@ -34,16 +34,16 @@ body: |
   ; EXPANDED-NEXT: {{  $}}
   ; EXPANDED-NEXT:   S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr12, 8, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr13, 9, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
-  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr14, 10, $vgpr0, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr12, 8, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr13, 9, $vgpr0
+  ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 $sgpr14, 10, $vgpr0
   ; EXPANDED-NEXT:   $vgpr0 = V_WRITELANE_B32 killed $sgpr15, 11, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
   ; EXPANDED-NEXT:   S_CBRANCH_SCC1 %bb.1, implicit undef $scc
   ; EXPANDED-NEXT: {{  $}}

diff  --git a/llvm/test/CodeGen/AMDGPU/vgpr-spill-scc-clobber.mir b/llvm/test/CodeGen/AMDGPU/vgpr-spill-scc-clobber.mir
index 9d3a0ce4b0c34..b036d4c9ec139 100644
--- a/llvm/test/CodeGen/AMDGPU/vgpr-spill-scc-clobber.mir
+++ b/llvm/test/CodeGen/AMDGPU/vgpr-spill-scc-clobber.mir
@@ -627,7 +627,7 @@ body:             |
   ; MUBUF-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
   ; MUBUF-NEXT:   $vgpr3 = V_MOV_B32_e32 8200, implicit $exec
   ; MUBUF-NEXT:   $vgpr0 = BUFFER_LOAD_DWORD_OFFEN $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.1, addrspace 5)
-  ; MUBUF-NEXT:   $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.1 + 4, addrspace 5)
+  ; MUBUF-NEXT:   $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.1 + 4, addrspace 5)
   ; MUBUF-NEXT:   $vgpr2 = BUFFER_LOAD_DWORD_OFFEN killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.1 + 8, addrspace 5)
   ; MUBUF-NEXT:   $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
   ; MUBUF-NEXT:   S_CBRANCH_SCC1 %bb.2, implicit $scc
@@ -690,7 +690,7 @@ body:             |
   ; VMEM-GFX8-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
   ; VMEM-GFX8-NEXT:   $vgpr3 = V_MOV_B32_e32 8200, implicit $exec
   ; VMEM-GFX8-NEXT:   $vgpr0 = BUFFER_LOAD_DWORD_OFFEN $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.1, addrspace 5)
-  ; VMEM-GFX8-NEXT:   $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.1 + 4, addrspace 5)
+  ; VMEM-GFX8-NEXT:   $vgpr1 = BUFFER_LOAD_DWORD_OFFEN $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (load (s32) from %stack.1 + 4, addrspace 5)
   ; VMEM-GFX8-NEXT:   $vgpr2 = BUFFER_LOAD_DWORD_OFFEN killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2 :: (load (s32) from %stack.1 + 8, addrspace 5)
   ; VMEM-GFX8-NEXT:   $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
   ; VMEM-GFX8-NEXT:   S_CBRANCH_SCC1 %bb.2, implicit $scc
@@ -960,7 +960,7 @@ body:             |
   ; MUBUF-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
   ; MUBUF-NEXT:   $vgpr3 = V_MOV_B32_e32 8200, implicit $exec
   ; MUBUF-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr0, $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.1, addrspace 5)
-  ; MUBUF-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr1, $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.1 + 4, addrspace 5)
+  ; MUBUF-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr1, $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.1 + 4, addrspace 5)
   ; MUBUF-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr2, killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.1 + 8, addrspace 5)
   ; MUBUF-NEXT:   $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
   ; MUBUF-NEXT:   S_CBRANCH_SCC1 %bb.2, implicit $scc
@@ -1023,7 +1023,7 @@ body:             |
   ; VMEM-GFX8-NEXT:   BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
   ; VMEM-GFX8-NEXT:   $vgpr3 = V_MOV_B32_e32 8200, implicit $exec
   ; VMEM-GFX8-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr0, $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.1, addrspace 5)
-  ; VMEM-GFX8-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr1, $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.1 + 4, addrspace 5)
+  ; VMEM-GFX8-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr1, $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.1 + 4, addrspace 5)
   ; VMEM-GFX8-NEXT:   BUFFER_STORE_DWORD_OFFEN $vgpr2, killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2 :: (store (s32) into %stack.1 + 8, addrspace 5)
   ; VMEM-GFX8-NEXT:   $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
   ; VMEM-GFX8-NEXT:   S_CBRANCH_SCC1 %bb.2, implicit $scc

diff  --git a/llvm/test/CodeGen/AMDGPU/vgpr-spill.mir b/llvm/test/CodeGen/AMDGPU/vgpr-spill.mir
index 99a9e2a9f71a1..c2badc5720f14 100644
--- a/llvm/test/CodeGen/AMDGPU/vgpr-spill.mir
+++ b/llvm/test/CodeGen/AMDGPU/vgpr-spill.mir
@@ -148,8 +148,8 @@ body:             |
     ; CHECK: liveins: $vgpr0_vgpr1_vgpr2_vgpr3
     ; CHECK-NEXT: {{  $}}
     ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0, addrspace 5)
-    ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 4, addrspace 5)
-    ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 8, addrspace 5)
+    ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+    ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
     ; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
     SI_SPILL_V128_SAVE killed $vgpr0_vgpr1_vgpr2_vgpr3, %stack.0, $sgpr32, 0, implicit $exec :: (store (s128) into %stack.0, addrspace 5)
 ...


        


More information about the llvm-commits mailing list